The Effect of Language On Economic Behavior

American Economic Review 2013, 103(2): 690–731
http://dx.doi.org/10.1257/aer.103.2.690
The Effect of Language on Economic Behavior:

Evidence from Savings Rates, Health Behaviors,
and Retirement Assets†
By M. Keith Chen*
Languages differ widely in the ways they encode time. I test the
hypothesis that the languages that grammatically associate the future
and the present, foster future-oriented behavior. This prediction
arises naturally when well-documented effects of language structure
are merged with models of intertemporal choice. Empirically, I
find that speakers of such languages: save more, retire with more
wealth, smoke less, practice safer sex, and are less obese. This
holds both across countries and within countries when comparing
demographically similar native households. The evidence does not
support the most obvious forms of common causation. I discuss
implications for theories of intertemporal choice. (JEL D14, D83,
E21, I12, J26, Z13)
Languages differ in whether or not they require speakers to grammatically mark

future events. For example, a German speaker predicting rain can naturally do so in
the present tense, saying: Morgen regnet es which translates to “It rains tomorrow.”
In contrast, English would require the use of a future marker like “will” or “is going
to,” as in: “It will rain tomorrow.”1 In this way, English requires speakers to encode
a distinction between present and future events, while German does not.2 Could this
characteristic of language influence speakers’ intertemporal choices?
In this paper I test a linguistic-savings hypothesis: that being required to speak
in a distinct way about future events leads speakers to take fewer future-oriented
* Yale University, 135 Prospect Street, New Haven, CT 06511 (e-mail: keith.chen@gmail.com). I am indebted
to Judy Chevalier, Östen Dahl, Ashwini Deo, Bob Frank, Shane Frederick, Emir Kamenica, Mark Liberman, Elisa
Long, John McWhorter, Sharon Oster, Ben Polak, Geoffrey Pullum, Frances Woolley, an anonymous referee and
seminar participants at Berkeley, Boston University, Harvard, Penn Linguistics, Stanford Economics, Stanford
Linguistics, UCLA, UCSD, Wharton, Yale Economics, and Yale Linguistics, for generous feedback and sugges-
tions. Special thanks are due to Nicole Palffy-Muhoray, who provided extensive feedback on multiple drafts of this
paper, and to Jane Bang and Ryan Caro for their research assistance. All errors are my own. I have no relevant or
financial interests related to this project to disclose.
†
To view additional materials and author disclosure statement(s), visit the article page at
http://dx.doi.org/10.1257/aer.103.2.690.
1
These are what linguists call periphrastic constructions, in which future events are marked by the addition of
auxiliary words.
2
In English, future reference is possible without future markers in certain contexts: specifically with scheduled
events or events resulting from law-like properties of the world. See Copley (2009) for details. In my analysis, I set
aside these cases because, as shown in Dahl (1985) and Dahl (2000, p. 311), “in many if not most languages, this
kind of sentence is treated in a way that does not mark it grammatically as having non-present time reference... even
for languages where future-time reference is otherwise highly grammaticalized.” In other words, how scheduled
events are treated does not reflect a language’s overall treatment of future reference.
690
VOL. 103 NO. 2 chen: The Effect of Language on Economic Behavior 691
actions. This hypothesis arises naturally if grammatically separating the future and
the present leads speakers to disassociate the future from the present. This would
make the future feel more distant, and since saving involves current costs for future
rewards, would make saving harder. On the other hand, some languages grammati-
cally equate the present and future. Those speakers would be more willing to save
for a future which appears closer. Put another way, I ask whether a habit of speech
which disassociates the future from the present, can cause people to devalue future
rewards.
The bulk of this paper investigates whether this prediction is borne out in savings
behavior. To do so, I first review the literature on what linguists call future-time
reference (FTR), which studies both when and how languages require speakers
to mark the timing of events. From this literature I adopt a future-time criterion
from typological linguistics, which separates languages into two broad catego-
ries: weak and strong-FTR. This criterion separates those languages that require
future events to be grammatically marked when making predictions (strong-
FTR languages, like English), from those that do not (weak-FTR languages, like
German).3 By analyzing text samples extracted from the web, I confirm that this
linguistic distinction captures a central tendency of how languages mark future
events, and that this distinction can be both generated and verified in automati-
cally collected data.
I then examine how these linguistic differences correlate with future-oriented
behaviors such as saving, exercising, abstaining from smoking, condom use, retire-
ment savings, and long-run health. I also attempt to determine if differences in
language cause these differences in behavior, or if non-linguistic traits that are coin-
cident with language explain these correlations. For example, most (but not all)4
Germanic languages are weak-FTR: could there also be a “Germanic” savings value
that is widely held by Germanic-language speakers but not caused by language?
While not conclusive, the evidence does not support the most obvious forms of
common causation.
Cross-country regressions show a strong correlation between weak-FTR languages
and future-oriented behavior, which do not attenuate with the inclusion of numerous
geographic, cultural, and institutional controls. Switching to within-country regres-
sions, I compare individuals with identical income, education, family structure, and
countries of birth, but who speak different languages. These regressions rely for
identification on a set of nine multilingual countries with both weak- and strong-
FTR populations.5 In these regressions, speakers of weak-FTR languages (with
little to no grammatical distinction between the present and future) appear more
3
Specifically, I adopt a criteria which distinguishes between languages which Dahl (2000, p. 325) calls “future-
less,” and those which are not. Dahl defines “futureless” languages as those which do not require “the obligatory use
[of grammaticalized future-time reference] in (main clause) prediction-based contexts.” In this framework, a predic-
tion is a statement about the future that has no intentional component. Predicting the weather would be a canonical
example. See Dahl (2000) and Thieroff (2000) for a discussion of the basis and areal properties of this distinction.
In the text of this paper, I adopt Thieroff’s more neutral language of “weak-FTR” for “futureless” languages, and
denote non-weak-FTR languages as “strong-FTR.” See Section IIIA for details on the EUROTYP criteria developed
by Dahl (2000), Appendix B for a discussion of how I apply this criteria to languages not covered by EUROTYP,
and the online Appendix for a complete list of coded languages.
4
Interestingly, English is a notable outlier among Germanic languages. I discuss this at length in Section II.
5
These countries are Belgium, Burkina Faso, Ethiopia, Estonia, The Democratic Republic of the Congo, Nigeria,
Malaysia, Singapore, and Switzerland.
692 THE AMERICAN ECONOMIC REVIEW april 2013
future-oriented in numerous monetary and non-monetary behaviors. Weak-FTR

speakers are 31 percent more likely to have saved in any given year, have accumu-
lated 39 percent more wealth by retirement, are 24 percent less likely to smoke, are
29 percent more likely to be physically active, and are 13 percent less likely to be
medically obese.
Similar to my cross-country regressions, the effect of language is not attenuated
by controls for cultural traits and values in these within-country regressions. Most
notably, several waves of the World Values Survey (WVS) asked respondents the
degree to which “savings and thrift is an important value.” I find that both future-
time reference and the degree to which a person reports valuing savings, predict
savings behavior. However, these effects are completely independent. This suggests
that the language effects I identify operate through a channel which is independent
of cultural attitudes toward savings.
Finally, I examine the effect that this differential propensity to save has on
national savings rates, both among the developed-country members of the
Organisation for Economic Co-operation and Development (OECD), and among
the larger set of WVS countries. Several interesting patterns emerge. First, a coun-
try’s language has a significant effect on that country’s aggregate savings rate.
Countries which speak weak-FTR languages save on average 6 percent more of
their GDP per year. This result is unaffected by the addition of life-cycle-savings
controls, holds in every major region of the world, and appears stable across time.
Parallel regressions using World Bank savings data show this same result among
developing nations, even with numerous controls for culture, values, institutions,
and legal origins.
The paper proceeds as follows. Section I reviews the linguistics literature on
future-time reference and details the ways it differs across languages. Section II
lays out my hypothesis and potential mechanisms, and discusses linguistics and
psychological studies that bear on these mechanisms. Section III details my
empirical methods and the data I use for estimation. Section IV presents condi-
tional correlations between a language’s FTR and its speakers’ future-oriented
behaviors. More detailed regressions investigate the degree to which these correla-
tions can be taken as evidence of causation. A final set of regressions investigates
the relationship between language and national savings rates. Section V discusses
several related literatures on the effect of language on thought, as well as the
implications of my findings for work in economics. Section VI discusses issues
surrounding the interpretation of my results before concluding.
I. Language and Future-Time Reference
Languages differ widely in both how and when they require speakers to signal that
they are talking about the future. For example, English primarily marks the future
with either “will” or forms of “be going to.” 6 In contrast, some languages mark
6
The English “will” is what linguists call a de-volitive future construction which descended from the Proto-
Germanic willan, or “want.” While “will” was already used for future-time reference in Old English, the mod-
ern German equivalent wollen, is not used for future-time reference in German. The English “be going to” is a
future events using a much larger and diverse set of constructions. For example,
Bittner (2005) documents that Kalaallisut (West Greenlandic), has at least 28 dis-
tinct constructions which mark future time:
“… nineteen verb-extending suffixes (sixteen transitivity preserving … ,

three transitive-deriving …), four verbal roots (one complex predicate
forming …), one noun-extending suffix … , one de-nominal verb-forming
suffix … and three mood inflections.”
More subtly, languages also differ in when they require speakers to specify the
timing of events, or when timing can be left unsaid. The linguist Roman Jakobson
explained this difference as: “Languages differ essentially in what they must convey
and not in what they may convey” (Jakobson and Halle 1956).
For example, if I wanted to explain to an English-speaking colleague why I can’t
attend a meeting later today, I could not say “I go to a seminar.” English grammar
would oblige me to say “I (will go, am going, have to go) to a seminar.” If, on the
other hand, I were speaking Mandarin, it would be quite natural for me to omit any
 ıˉn  g ji   aˇ ngzuò (I go listen seminar):
marker of future time and say W   oˇ qù t   
(1) W   oˇ  qù t   

 ıˉn  g ji   aˇ ngzuò
I go.prs listen seminar
“I am going to listen to a seminar”
with no reference to future time, since the context leaves little room for
misunderstanding.7
In this way, English forces its speakers to habitually divide time between the pres-
ent and future in a way that Mandarin (which has no tenses) does not. Of course,
this does not mean that Mandarin speakers are unable (or even less able) to under-
stand the difference between the present and future, only that they are not required
to attend to it every time they speak. This difference, in the obligatory marking of
future events is a central characteristic of the weak- versus strong-FTR classification
(Thieroff 2000), and is the difference between languages I exploit in my study of
savings behaviors.
These differences between languages are surprisingly widespread, and occur not
only between neighboring countries in the same region, but sometimes occur within
multilingual countries. For example, European languages range from a tendency to
rarely distinguish present and future time (like Finnish) to languages like French,
which have separate and obligatory “future” forms of verbs.8 A Finnish speaker
de-andative construction that developed in the seventeenth century, and is found in English, Dutch, French, Spanish,
and Portuguese, but not in German. In English, neither construction is purely a tense marker, but instead mark
different temporal and modal properties which give rise to future-time reference in certain contexts (“going to” is
prospective aspect, while “will” can be a modal auxiliary).
7
In this and all subsequent examples, I follow the Leipzig glossing rules, where fut and prs indicate future and
present morphemes. See Croft (2003) for details.
8
Languages where verbs have distinct future forms are said to have an “inflectional” future. In Europe, this
includes the romance languages (except Romanian), and most Slavic and Semitic languages. See Dahl (1985)
for source data on inflectional futures in Europe, and Dahl and Velupillai (2011) for a broad survey of inflectional
futures around the world.
for example, would say both Tänään on kylmää (today is cold) and Huomenna on
kylmää (tomorrow is cold) using the unmarked verb on:
(2) a. Tänään on kylmää
Today be.prs cold

“It is cold today”
b. Huomenna on kylmää
Tomorrow be.prs cold
“It will be cold tomorrow”
while French speakers would switch from Il fait froid aujourd’hui (it is cold today),
to Il fera froid demain (it will-be cold tomorrow):
(3) a. Il fait froid aujourd’hui

It do/make.prs cold today
“It is cold today”
b. Il fera froid demain

It do/make.fut cold tomorrow
English is a notable outlier in Europe; in all other Germanic languages grammatical

future-time reference is optional when making predictions that have no intentional
component.9 That is, while a German speaker predicting rain or forecasting a freeze
could say Morgen regnet es, or Morgen ist es kalt (both in the present tense):
(4) a. Morgen regnet es

Tomorrow rain.prs it
“It will rain tomorrow”
b. Morgen ist es kalt

Tomorrow is.prs it cold
an English speaker would have to grammatically mark future time (it will rain
tomorrow, and it will be cold tomorrow).10 Later, I will exploit the fact that weak-
and strong-FTR languages often coexist within the native languages of the same
9
This observation that German and English differ dramatically in obligatory GFTM is not new: Comrie (1985)
cites English and German as exemplars of strong- and weak-FTR languages. For a detailed analysis of this differ-
ence between English and German see Copley (2009). Copley demonstrates that in English, “futurates” (sentences
about future events with no FTR) can only be used to convey information about planned/scheduled/habitual events,
or events which arise from law-like properties of the world. This restriction is not present in German, and futurates
are common in German speech and writing.
10
Thieroff (2000) documents what Dahl (2000) calls a “futureless area” in Northern and Central Europe, includ-
ing most Finno-Ugric and all Germanic languages except English.
country, helping me isolate linguistic effects from confounds that vary on the coun-
try level (such as taxes, institutions, or capital markets).
II. A Linguistic-Savings Hypothesis
My linguist-savings hypothesis stems naturally from two different types of mech-

anisms, one concerning a linguisticly-induced bias in time perception, and one con-
cerning the precision of beliefs about time. To illustrate these mechanisms, consider
a simple savings problem. Suppose a decision maker is deciding whether or not to
pay cost C now, in exchange for reward R > C at some time in the future. Suppose
she is uncertain about when reward R will materialize, and holds beliefs with dis-
tribution F(t). If the decision maker discounts future rewards at rate δ, then she will
prefer to save/invest if and only if
(5) C < ∫ e−δ
t Rd F(t).
A. Mechanism One: Obligatory Distinctions Bias Beliefs
The first way that language may affect future choices is by changing how distant
future events feel. For example, it seems plausible that speaking about future events
as if they were happening now (in the present tense), would lead weak-FTR speak-
ers to perceive future events as less distant. Indeed, several literary and rhetorical
techniques appear to depend on this for their effectiveness.
For example, speakers often narrate past events in the present tense, with the goal
of making those events feel more vivid and immediate (linguists call this the histori-
cal present). Writers have consciously used this tense-shifting strategy since at least
the first century, when in what is often considered the first treatise on writing style,
Longinus wrote:
“If you introduce things which are past as present and now taking
place, you will make your story no longer a narration, but an actuality.”
(Longinus, On the Sublime, first-century AD)
For example, the historian Peter Rodman writes:
“There is a famous story of President Abraham Lincoln, taking a vote in

a cabinet meeting on whether to sign the Emancipation Proclamation. All
his cabinet secretaries vote nay, whereupon Lincoln raises his right hand
and declares: “The ayes have it.’” (emphasis mine. Rodman, Presidential
Command 2009, p. 3).
Common across languages and used in both writing and conversation, linguists have
traditionally thought that narrating the past in the present make a story more vivid
and immediate by “moving past events out of their original time frame and into the
moment of speaking” (Schiffrin 1981).
Indeed, this technique is widespread enough to elicit scorn from critics who con-
sider it manipulative. The Man Booker Prize judge Philip Hensher attributes the
routine use of the present tense to: “a thousand low-level creative writing tutors,
clinging to the belief that you can ‘make your writing more vivid’ by turning to the
present tense.” Author and critic Philip Pullman (while acknowledging skillful use
by Brontë and Dickens), agrees, complaining: “if every sound you emit is a scream,
a scream has no expressive value … I feel claustrophobic, always pressed up against
the immediate.”
Similarly, jokes are almost always told in the present tense, as in “a man walks
(not walked) into a bar.” This is often attributed to a need for immediacy and sur-
prise in successful humor. If talking about temporally distant events in the present
tense makes them seem more immediate, languages which force this way of speak-
ing may make future events seem closer.
A similar linguist construction flips this logic to arrive at the same effect. In what
some linguists call distancing, distant (past or future) tenses are used instead of the
present tense to convey that a current event is distant in some other way (Dancygier
and Sweetser 2009).11 For example, something that is distant from current reality
(unlikely or impossible), is often spoken about in the past tense, as in:
“I wish I had a car (right now).”
which conveys that the person does not have a car at present. Note that the past-tense
“had” is not meant to describe a past state, which would require an auxiliary “had.”
“I wish I had had a car (when I was a student).”
Iatridou (2000) notes that this use of the past tense to convey distance from reality
occurs in both wishes (like above) and also in if-clauses:
“If I had a car (I would give you a ride),” “If he ran the school… .”
These non-past uses of past tenses are widespread across languages (James 1982),
and while less common, future tenses are also used (Mezhevich 2008). If speaking
about current events in a distant tense makes them seem distant, languages which
oblige speakers to use a future (distant) tense may make future events seem more
distant.
In the context of my simple model, this could be represented either through the
discount rate, or through beliefs. That is, we could imagine that weak- and strong-
FTR languages lead speakers hold rates δW < δS , which would immediately trans-
late into different willingness to save:
(6)  < δS, then ∫ e−

if δW δ W t Rd F(t) > ∫ e−
δ S t Rd F(t).
(t) and
Alternatively, we could represent such an effect as shifting the beliefs FW
FS(t) that weak- and strong-FTR speakers hold. If weak-FTR speakers perceive the
11
This use of the simple past to convey epistemic distance is closely related to several well-studied linguis-
tic phenomena, including hypotheticals (see James 1982), counterfactuals (see Iatridou 2000), conditionals (see
Dancygier and Sweetser 2009), and most broadly, the irrealis and subjunctive moods.
S(t) would first-order stochastically dominate F

future as closer, then F W(t). It is easy
to see how this would affect the decision to save:

(7) if ∀t, FW(t) ≥ FS(t), then   ∫   e  −δt
RdFW
(t) ≥  ∫   e  −δt
RdFS(t).
B. Mechanism Two: Linguistic Distinctions Lead to More Precise Beliefs
The second way that language may affect future choices is by leading speakers
to have more or less precise beliefs about the timing of future rewards. Languages
with more grammatical time marking would lead speakers to hold more precise
beliefs about the timing of events if either: marking time requires increased atten-
tion to time, or if these markers are encoded in memory. While no studies (to my
knowledge) have directly examined the effects of how a language treats time, a large
literature has found that language with more precise “basic color terms”12 cause
their speakers to hold more precise color beliefs.
Summarizing this literature, MacLaury et al. (1992) notes that languages around
the world possess anywhere from 2 to 11 basic color terms. So for example, while
almost all languages distinguish between black, white, and red, several languages
name all of yellow, green, and blue with one basic color term, and many languages
do not have a basic word for purple, pink, orange, or grey. In one of the first stud-
ies examining the cognitive correlates of these differences, Brown and Lenneberg
(1954) find that Zuñi speakers (who lack a lexical distinction between orange and
yellow) have trouble remembering nuanced differences between orange/yellow
colors.13
More recent studies have confirmed the direct role of language in these findings.
Russian makes an obligatory distinction between light blue (goluboy) and dark blue
(siniy). Winawer et al. (2007) finds that Russian speakers do better than English
speakers in distinguishing blues when the two colors span the goluboy/siniy border
(but not when then do not), and that these differences are eliminated when subjects
must simultaneously perform a verbal (but not a spatial) distractor task. Further
implicating language in this differential precision, Franklin et al. (2008) finds that
this difference holds for adults, but not for pre-linguistic infants.
If this linguistic-precision effect is also true for time perception, then strong-FTR
speakers will be less willing to save. To see this, assume strong-FTR speakers (who
must separate the future and present) hold more precise beliefs about the timing of
reward R than speakers of weak-FTR languages. More concretely, if F (t) and FS (t)
W
are the beliefs of weak-FTR and strong-FTR language speakers, then we might
expect FW (t) to be a mean-preserving spread of FS (t). Proposition 1 establishes that
a decision maker with beliefs FW (t) will value future rewards more than one who
holds beliefs FS (t).
 
Proposition 1: If FW (t) is a mean-preserving spread of FS (t), then ∫    e
  −δt  Rd FW (t)
  −δt
> ∫    e
    Rd FS (t).
12
MacLaury et al. (1992) defines “basic color terms” as: “the simplest forms of broadest meaning that most
speakers of a language will routinely apply to colors in any context.”
13
The Zuñi (one of the Pueblo peoples), are a Native American tribe that live primarily in western New Mexico.
Proof:
(t) is a mean-preserving spread of FS (t), then FS (t) second-order
Note that if FW
stochastically dominates F (t). Also note that for any discount rate δ > 0, e −δt
W is a
 
strictly-convex function. Therefore ∫   e  −δt
  Rd FW (t) > ∫  e  −δt
  Rd FS (t).
In other words, if more finely partitioning events in time leads to more precise
beliefs, weak-FTR speakers will be more willing to save than their strong-FTR
counterparts. Intuitively, since discounting implies that the value of future rewards
is a strictly-convex function of time, uncertainty about the timing of future payoffs
makes saving more attractive. Experimentally, Redelmeier and Heller (1993) find
this risk-seeking response to timing uncertainty, which is also commonly observed
in animal studies (see Kacelnik and Bateson 1996). Note that the exponential dis-
count function I specified in equation (5) is not critical: every widely studied theory
of discounting is strictly convex, and would produce the same result.14
This would have the same effect on savings as mechanism one: people who speak
weak-FTR languages (who speak the future and present identically) would save,
exercise, and plan more, and spend, smoke, and over-consume less. I will now pres-
ent a set of empirical findings which test this hypothesis, then return to a more gen-
eral discussion of language and cognition.
III. Data and Methods
A. Coding Languages
In all of the regressions to follow, the independent variable of main interest is

Strong FTR (strong future-time reference), a criterion I did not develop but adopt
from the European Science Foundation’s Typology of Languages in Europe
(EUROTYP) project.15 Summarizing the findings of the EUROTYP project, Dahl
(2000) describes a set of languages he calls “futureless” as those which do not
require “obligatory [FTR] use in (main clause) prediction-based contexts.” In this
paper, I adopt the more neutral language of “weak-FTR” for “futureless” languages,
and call non-weak-FTR languages “strong-FTR.” That is, English is a strong-FTR
language because marking future-time is obligatory in all but a small set of circum-
stances, even when making predictions that have no intentional component (e.g.,
“tomorrow it will be warm,” which the speaker is not promising to cause.)
This distinction between intentions we may have about things under our control,
versus mere predictions is a central distinction in the typology of FTR. Thieroff
(2000, p. 288) notes that at least in Europe, this distinction maps more generally onto
whether future events can be left unmarked (i.e., discussed in the present tense); that
is, in weak-FTR languages in general, “future time reference can be referred to with
14
See Frederick, Loewenstein, and O’Donoghue (2002) for a review of both models and evidence on discount-
ing behavior.
15
Future-time reference was a focal area of the EUROTYP Theme Group on Tense and Aspect, which studied
the typological and areal distribution of grammaticalized future-time reference. The idea for EUROTYP was devel-
oped at a European Science Foundation conference (Rome, January 1988). At those meetings, it was established
that a cross-linguistic study of the tense and aspect systems of European Languages would form one of EUROTYP’s
nine focus areas. The resulting working group summarized their findings in an 846 page volume on Tense and
Aspect, edited by Östen Dahl (2000).
unmarked form (the present), in other words, in general the future is not obligatory.”
Dahl also finds that weak-FTR corresponds strongly with a language’s general ten-
dency to require FTR, and that “whether FTR is overtly and obligatorily marked in
prediction-based sentences can be used as one of the major criteria for whether it is
grammaticalized in a language or not” (Dahl 2000, p. 310). These analyses motivate
my decision to use weak-FTR as a proxy for the general treatment of future time in
a language.
Most regressions in this paper cover languages directly analyzed by the EUROTYP
Theme Group. In those regressions, strong-FTR languages are the exact complement
of what Dahl calls “futureless” and Thieroff (2000) calls “weakly-grammaticalized
future” languages. Some regressions analyze the World-Values Survey, whose
participants speak many non-European languages not analyzed by either Dahl or
Thieroff. To extend their characterization to this broader set, I rely on several other
cross-linguistic analyses, (most notably Bybee, Perkins, and Pagliuca 1994; Cyffer,
Ebermann, and Ziegelmeyer 2009; Dahl 1985; Dahl and Kós-Dienes 1984; and
Nurse 2008), and on individual grammars for languages that are extensively spoken
in the WVS but not covered by these broader analyses. A more detailed discussion
of coding languages is found in Appendix B, and a large table of all languages
included in this study and their coding is included in the online Appendix.
B. Alternative Codings
While in this paper I focus for simplicity on weak- versus strong-FTR languages,
there are several related criteria that may be important. A weaker criterion might
be the presence of any grammatical marking of future events in a language, even
if infrequently used. This would include both inflectional markers (like the future-
indicating suffixes in Romance languages) or periphrastic markers (like the English
auxiliary “will.”) Mandarin, Finnish, and Estonian are examples of languages that
lack either type of future markers.16 A stronger criterion might be the presence of
an inflectional future tense, which would include most Romance languages, but
exclude English. These alternative criterion would satisfy
?
(8) Any Gr FTR ⊃ Weak FTR ⊃ Strong FTR ⊃   Inflectional FTR,
with the first and second inclusions being logically necessary, and the third repre-
senting a typological regularity for which I do not have a counterexample.17
For simplicity and transparency, in this paper I have adopted the main criterion
advocated by the EUROTYP working group for “futureless” languages, which
corresponds to the second inclusion. An additional reason for this choice is that,
as both Dahl and Thieroff note, in the EUROTYP data weak-FTR languages are
those in which “the future is not obligatory in sentences with future-time reference”
(Thieroff 2000, p. 288). Since this is the characteristic of languages (a more or less
16
Dahl (2000) writes that Finnish and Estonian stand out in Europe as “extreme examples of languages with
no systematic marking of future time reference (although this does not imply a total absence of devices that show
future time reference).”
17
More specifically, the languages which satisfy these criteria are nested sets. That is: {X | X has any grammati-
cal FTR} ⊃ {X | X has at least weak FTR} ⊃ {X | X has strong FTR} ⊃ {X | X has inflectional FTR}.
granular obligatory d iscretization of future time) that is central to the mechanism I

propose, differences between weak- and strong-FTR languages seem the most direct
test of my hypothesis.18
Online-Text Based Codings.—To my knowledge, the EUROTYP project is the

most extensive typological research program to study the cross-linguistic grammati-
calization of FTR.19 Nevertheless, it may be important to assess whether the linguis-
tic distinction I adopt from them can be validated independent of expert judgement.
To do this, I attempt to form a measure of FTR strength based on word-frequency
analysis of text retrieved from the web, and investigate how much this measure cor-
relates with weak- versus strong-FTR.
As a basis for an online measure, I scrape the web for full-sentence weather fore-
casts. Using weather forecasts has the advantage of comparing relatively controlled
sets of texts about future events.20 An important limitation of this approach is its
restriction to languages which are widespread on the Internet. As of the writing of
this paper, this results in a set of 39 coded languages.
Details and the results of this exercise are summarized in Appendix B, which
reports two measures of how frequently a language grammatically marks future
time. “Verb ratio” counts the number of verbs which are grammatically future-
marked, divided by the total number of future-referring verbs. In other words, in
online weather forecasts in a language, what share of verbs about future weather
are marked as future-referring? Similarly, “sentence ratio” asks: what share of sen-
tences regarding future weather contain a grammatical future marker? 21
Unsurprisingly, across languages these measures are highly correlated (0.992).
Both are also highly rank-correlated with the EUROTYP criterion (1.000). While I
cannot know how well these results would extend to languages which are not well
represented online, in general these results suggest that the EUROTYP criterion
measures an objective central tendency of a language’s FTR strength. For simplicity
then, all results presented in this paper report the average effect of moving from a
weak- to strong-FTR language.22
18
As a robustness check, it is possible to include all three inclusions as nested effects in the broader cross-coun-
try savings regressions in this paper. While these regressions do not have enough statistical power to disentangle
all three effects, results suggest increasingly strong effects as you move from weaker to stronger criteria, with joint
statistical significance similar to that of the binary weak-vs-strong distinction. Please see online Appendix Table 4
for the results of these regressions.
19
Sponsored by the European Science Foundation, EUROTYP involved about a hundred linguists over five
years (1990–1994), and its report on Tense and Aspect runs over 800 pages.
20
Another advantage is that weather forecasts are likely to reliably be restricted to prediction-based FTR mark-
ers, since weather forecasters do not generally believe they can affect the weather. This restriction comes with the
disadvantage that for many languages, prediction-based FTR is a small share of overall FTR strategies. However,
focusing on a consistent source of future-time predictions eliminates the worry that the relative proportions of FTR
strategies represented online may vary by language. In addition, using prediction-based FTR as a proxy for general
FTR tendencies is supported by the Thieroff (2000) observation that (at least in the EUROTYP data), the tendency
to mark FTR in prediction-based contexts maps more generally onto the obligatory FTR marking.
21
In some languages (Arabic, for example), a sentence with multiple verbs will often mark only the first as
future-regarding. Grammatical differences across languages like these produce variation between verb and sentence
ratios.
22
While the set of languages I can code this way is limited to those which are well represented on the searchable
Internet (39 as of the writing of this paper), it is extensive enough that both the OECD and SHARE results I report
can be run using either ratio instead of the binary strong-vs-weak FTR criteria. Both ratios produce results that are
nearly identical to the results I report, both quantitatively and statistically. Please see online Appendix Tables 1, 2,
and 3 for the results of these regressions.
C. Savings Regressions in the WVS
My first set of regressions are run on individuals in the World Values Survey
(WVS), a global survey of world cultures and values (World Values Survey 2009).
Although five waves of the WVS are available, I study only the last three, which ran
from 1994 to 2007. In these (but not earlier) waves, participants were asked what
language they normally speak at home, which I use as a proxy for the language most
likely to structure their thought. This allows me to study individuals across a set of
76 countries for which language data are available. For my purposes, this sample
has the nice feature that not only permits a broad set of countries with which to do
cross-country comparisons, it also includes a number of countries with sufficient
within-country linguistic diversity to permit within-country comparisons.
I estimate fixed-effect logit models of an individual’s propensity to save (versus
not save) in the current year, regressed on the FTR strength of that individual’s
language and a rich set of fixed-effects for country and individual characteristics.23
These fixed effects control for a person’s: country of residence, income decile within
that country, marital status (with six different classifications), sex, education (with
eight different classifications), age (in ten-year bins), number of children, survey
wave, and religion (from a set of 74) all interacted (for a total of 1.4 billion catego-
ries). Effectively, this analysis matches an individual with others who are identical
on every dimension listed above, but who speak a different language. It then asks
within these groups of otherwise identical individuals, do those who speak strong-
FTR languages behave differently than those who speak weak-FTR languages? In
addition, immigrants are excluded from this analysis so as to avoid conflating dif-
ferences in a household’s primary language with differences between natives and
immigrants.
The WVS also allows me to study the interaction between the effect of language
on savings behavior, and several beliefs and values questions asked of participants.
This allows me to examine to what degree the measured effect of language on sav-
ings behavior is attenuated by such things as how much a person reports trusting
other people, or how much they report that saving is an important cultural value. To
a limited extent, this allows me to investigate whether language acts as a marker of
deep cultural values that drive savings, or whether language itself has a direct effect
on savings behavior.
D. Retirement Assets in the SHARE
The WVS focusses on current savings behavior and beliefs. The second dataset
I analyze is the Survey of Health, Ageing, and Retirement in Europe (SHARE),
which is a panel survey that measures the socioeconomic status and health of retired
households in 13 European countries (Börsch-Supan and Jürges 2005). This allows
me to complement my earlier analysis of the WVS with an analysis of past savings
behavior (as reflected in accumulated wealth).
23
I use Chamberlain’s (1980) fixed-effect (or conditional) logit model to estimate these regressions, since I have
very few observations within each group defined by my fixed effects. The Chamberlain model solves the resulting
incidental-parameters problem.
Using the SHARE, I estimate several OLS models of total net household retire-
ment assets regressed on a household’s language.24 The SHARE attempts a compre-
hensive measure of all assets a household has, including income, private and public
benefit payments, and all forms of assets (stocks, bonds, housing, etc.) The richest
of these regressions includes fixed effects for a household’s: country of residence
(13), income decile within that country, marital status (with six different classifica-
tions), sex, education (with eight different classifications), age, number of children,
and survey wave (2004 and 2006), all interacted for a total of 2.7 million categories.
Like my regressions in the WVS, my analysis of the SHARE allows me to control for
many confounds by focusing on the effects of within-country language variation.
E. Health Behaviors in the SHARE and DHS
In addition to retirement assets, the SHARE also collects data on health behav-
iors such as smoking and exercise, as well as multiple measures of long-run phys-
ical-health: body-mass-index, walking speed (as measured by a walking test), grip
strength (as measured by a dynamometer), and respiratory health (peak expira-
tory air flow). This allows me to run similar fixed-effect regressions as my savings
results, but with a broader set of future-regarding behaviors.
Complementing these results, I run similar health regressions in data from the
MEASURE DHS project, which conducts demographic and health surveys in devel-
oping countries on behalf of USAID. These surveys collect nationally representative
data on fertility, family planning, and health behaviors in a large sample of develop-
ing countries. Together, these results allow me to investigate whether the savings
behavior results I study extend to health behaviors such as exercise, smoking, family
planning, and condom use, both in developed and developing countries.
F. National Savings Rates
Finally, I study the relationship between language and savings in a cross-country

framework, using national accounts data from both the OECD (1970 to present)
and World Bank data merged with the WVS. The OECD data are collected and
harmonized across all 34 member countries as well as for the Russian Federation.25
Details on the exact construction of each measure can be found in Appendix A.
While harmonized World Bank and WVS data do not go back as far as those from
the OECD, those data allow me to complement my OECD analysis with regressions
over a much broader set of countries at different levels of development, and which
include country-level controls for culture, attitudes, and beliefs.
These regressions ask whether the FTR structure of a country’s language appears
to affect national savings. The form of the national savings equation is a simple lin-
ear relation that follows closely from life-cycle savings theory (see Modigliani 1986
for a review). Essentially, I regress national-savings rates on the level and growth
24
Unfortunately, the SHARE does not record what language households speak at home. Instead, I exploit the
fact that the survey instrument is offered in multiple languages; households can choose to take the survey in any of
the national languages of their country. I use this choice as a proxy for their primary language.
25
I include the Russian Federation in this analysis because, as of the writing of this paper, they were in the pro-
cess of joining the OECD, and were included in the harmonized OECD data.
rate of GDP as well as a number of other country demographics. To this regression

I add a weighted measure of the FTR strength of that country’s languages. This is
simply the FTR strength of each of that country’s major languages, weighted by
the percent of the country’s population reports speaking those languages.26 This
language measure does not vary by year: these regressions test if the unexplained
components of national savings vary cross-sectionally with a country’s language,
and do not try to identify off of demographic shifts within a country across time.
IV. Results
The results that follow use data from five main sources of data: the WVS, the
WDI, the SHARE, the DHS, and the OECD. Please see Appendix A for a detailed
description of each dataset, as well methodologies, definitions, means, and standard
deviations of all variables of interest.
A. Language and Savings in the World Values Survey
My first set of regressions examines the savings behavior of individuals. I examine

this behavior using cross-country regressions where the dependant variable saveitis
an individual reporting having net saved this year.27 I estimate the equation
exp(zit)
(9) Pr (saveit) =  _
      ,
1 + exp(zit)
where
zit  = β  0  + β  1  StrongFTR  + β  2 Xit  + β  3 Xt  + β  4 F  ex
it    + β  5 F  t  .
c
In equation (9), the main variable of interest StrongFTR is a binary-coded charac-

teristic of the language that the individual speaks at home. Xitare characteristics of
individual i at time t, such as their employment status or self-reported beliefs about
trust and savings. X t are characteristics of a country at time t, such as their legal
system, economy, and country-level averages of variables like trust. F it  is a set of
  ex
fixed effects that can be taken as exogenous, these are non-choice variables such as
age and sex. F  ct  is a set of continent fixed effects. Empirical estimates of equation (9)
are presented in Tables 1 and 2; all coefficients are reported as odds ratios, where
the effect is 1.
Regression 1 controls only for F  ex it  , (non-choice variables age and sex), so as to
summarize the average difference in the propensity to save between strong- and
weak-FTR individuals. The coefficient of 0.460 can be interpreted as strong-FTR
families saving only 46 percent as often as weak-FTR families.
Regression 2 adds the origin of a country’s legal system and the log of its
PCGDP, mirroring the cross-country regressions La Porta, Lopez-De-Silanes, and
26
These relative language shares were obtained for each country from their national census taken closest to the
year 2000.
27
See Appendix A for the exact wording of this and other questions in the WVS.
Table 1—An Individual Saved This Year (WVS, Cross-Country Analysis)
Saved
(1) (2) (3) (4) (5) (6)
Strong FTR 0.460 0.441 0.443 0.449 0.456 0.471
[0.069]*** [0.071]*** [0.072]*** [0.073]*** [0.074]*** [0.073]***
French legal origin† 0.473 0.579 0.595 0.579 0.59
[0.065]*** [0.110]*** [0.129]** [0.128]** [0.128]**
German legal origin† 0.406 0.441 0.451 0.435 0.449
[0.084]*** [0.096]*** [0.134]*** [0.147]** [0.150]**
Scandinavian legal origin† 0.616 0.665 0.670 0.643 0.683
[0.192] [0.215] [0.324] [0.355] [0.371]
log per capita GDP† 1.164 1.154 1.171 1.169 1.168
[0.057]*** [0.057]*** [0.093]** [0.098] [0.100]
PCGDPt−1 / PCGDP  †t  0.597 0.588 0.600 0.580
[0.672] [0.676] [0.675] [0.655]
Unemployed 0.532 0.529 0.529 0.529
[0.035]*** [0.038]*** [0.036]*** [0.036]***
Real interest rate† 0.996 0.997 0.996 0.997
[0.003] [0.003] [0.003] [0.003]
WDI legal-rights index† 1.045 1.044 1.042 1.038
[0.035] [0.044] [0.048] [0.049]
Trust 1.229 1.229
[0.041]*** [0.041]***
Family is important 0.886 0.886
[0.024]*** [0.024]***
Trust† 0.906 0.953
(country average) [0.493] [0.535]
Family is important† 1.602 1.629
(country average) [1.500] [1.557]
Language share 0.911
[0.149]
FTR share 0.775
[0.339]
Fixed effects:
Age  ×  sex Yes Yes Yes Yes Yes Yes
Continent No No No Yes Yes Yes
Observations 152,056 149,350 140,498 140,498 134,535 134,535
Notes: Regressions are logistic regressions with coefficients reported as odds ratios. Immigrants are excluded from
all regressions. Variables with a (†) following their name vary only at the country level. Robust standard errors are
reported in brackets; all regressions are clustered at the country level.
*** Significant at the 1 percent level.
** Significant at the 5 percent level.
Shleifer (2008) use to study the effect of legal origins. Regression 3 adds a richer
set of controls than typically included in those regressions, including the growth
rate of PCGDP, unemployment, real interest rates, and the WDI legal-rights index.
Regression 4 goes further, adding continent fixed effects. If anything, the inclusion
of these controls increases the measured effect of language.
Regression 5 adds the two most studied variables in the large literature on social
capital as additional controls, both at the household level, and as their country-
level averages. “ Trust” measures whether an individual thinks “most people can be
trusted.” This measure has a large and marginally significant effect on the propensity
of an individual to save; individuals who think others are generally trustworthy are
on average 23 percent more likely to have saved this year. “Family” measures how
important a respondent says that family is to them (with 1 being “very” and 4 being
“not at all”.) People who report valuing family save significantly more than those
who do not. Both of these effects appear to be largely independent of the effect of
language.
Regression 6 adds controls for the share of a country which speaks a household’s
language, and what share speak a language with the same FTR level. These results
demonstrate that the effect of language is not driven by speaking either minor-
ity languages or FTR structures. Table 2 presents the coefficient on strong-FTR
when this final regression (regression 6) is run separately by continent and level of
development.
These regressions display coefficients less than one in every major region of the
world and at every level of PCGDP, consistent with my main findings. Estimates
from these first sets of regressions suggest that a language’s FTR is an important
predictor of savings behavior. This effect is large (larger than that of other widely-
studied variables), does not attenuate despite the inclusion of an aggressive set of
controls, and appears to hold across both geographical regions and levels of devel-
opment. Nevertheless, these regressions are fundamentally cross-country, and may
omit important differences between countries not captured by standard controls.
To attempt to account for this possibility, my next set of regressions include
both country fixed effects and comprehensive household-level controls, compar-
ing demographically similar households born and living in the same country. These
within-country regressions rely on the fact that many countries are multilingual, and
contain sets of extremely similar native families who live in close geographic prox-
imity, but who speak different languages.
These regressions are carried out using fixed-effect (or conditional) logistic anal-
ysis, where the dependant variable saveitis an individual reporting having saved in
net this year.28 I estimate the equation
exp(zit)
(10)   Pr   (  saveit) =  _

      ,
1 + exp(zit)
where
zit = β1  StrongFTR + β2 Xit + β3 F  ex
it   × 
F  en
it   × 
F  ct  .
In equation (10), the F variables are sets of fixed effects that are jointly interacted
to form groups for the basis of analysis: the conditional-likelihood function is cal-
culated relative to these groups. That is, individuals are compared only with others
who are identical on every F variable. F it  is a set of fixed effects that can be taken
  ex
as exogenous, these are non-choice variables such as age and sex. F  en it  is a set of
fixed effects that are likely endogenous to an individual’s discount rate, such as
income, education, and family structure. F   ct  is a set of country-wave fixed effects. In
using these extensive fixed effects to compare like families, this estimation strategy
mirrors that of Poterba, Venti, and Wise (1995) and the international comparisons
28
See Chamberlain (1980) for details on conditional-logistic analysis, and Appendix A for the exact wording of
this and other questions in the WVS.
Table 2—WVS Cross-Country Analysis by Continent and PCGDP
Strong FTR
Regression restricted by continent: Coefficient SE Observations
Africa 0.596 [0.095]*** 28,262
Asia 0.519 [0.104]*** 30,198
Europe 0.581 [0.135]** 45,502
Americas 0.713 [0.148] 26,854
Regression restricted by PCGDP:

PCGDP  ≤  1, 000 0.285 [0.093]*** 38,271
1, 000  <  PCGDP  ≤  5, 000 0.743 [0.172] 56,403
5, 000  <  PCGDP  ≤  25, 000 0.680 [0.163] 29,732
PCGDP  >  25, 000 0.422 [0.018]*** 10,123
Notes: Coefficients (reported as odds ratios) are from logistic regressions with the same specification as regression
6 in Table 1, but restricted by continent or level of per capita GDP. Robust standard errors are reported in brackets;
all regressions are clustered at the country level.
of household savings in Poterba (1994). Empirical estimates of equation (10) are

presented in Tables 3, 4, and 5; all coefficients are reported as odds ratios.
Included again for the sake of comparison, regression 1 controls only for
it  , (
F  ex non-choice variables age and sex), so as to summarize the average difference
in the propensity to save between strong- and weak-FTR individuals. Regressions 2
and 3 add fully-interacted fixed effects for country, time, income, and education.
On top of these, regressions 4 through 6 include controls for family structure.
Regression 4 can be interpreted as demonstrating that even when comparing only
individuals that are identical on every dimension discussed above, individuals who
speak a language with strong-FTR are roughly 30 percent less likely to report hav-
ing saved this year. This effect is nearly as large as being unemployed (31 percent).
As before, regression 5 adds “ Trust,” which has a marginally significant effect on
the propensity of an individual to save. Individuals who think others are generally
trustworthy are on average 8 percent more likely to have saved this year. “Family”
measures how important a respondent says that family is to them (with 1 being
“very” and 4 being “not at all”). People who report valuing family save significantly
more than those who do not. Both of these effects appear to be largely independent
of the effect of language. Indeed, by comparing regressions 4 and 5 we see that the
inclusion of “ Trust” and “Family,” if anything, increases the measured effect of
language.
Regression 6 adds a variable intended to measure saving as an important cultural
value. Specifically, this question asks whether “thrift and saving money” is a value
which is important to teach children.29 Unsurprisingly, individuals who report that
saving money is important are more likely to save. However, this effect is both
smaller than the effect of language (11 percent versus 30 percent), and does not
meaningfully attenuate the effect of language on savings behavior. This can be seen
by comparing the coefficients on Strong FTR between regressions 5 and 6. Indeed,
29
See Appendix A for the full wording of these questions in the WVS.
Table 3—An Individual Saved This Year (WVS, Within-Country Analysis)
Saved
(1) (2) (3) (4) (5) (6)
Strong FTR 0.460 0.718 0.722 0.700 0.691 0.693
[0.069]*** [0.113]** [0.115]** [0.103]** [0.090]*** [0.092]***
Unemployed 0.677 0.694 0.688 0.689
[0.031]*** [0.044]*** [0.044]*** [0.044]***
Trust 1.083 1.084
[0.045] [0.045]
Family is important 0.952 0.953
[0.057] [0.057]
Saving is important 1.111
(to teach children) [0.044]***
Fixed effects:
Age  ×  sex Yes Yes Yes Yes Yes Yes
Country  ×  wave No Yes Yes Yes Yes Yes
Income  ×  edu No Yes Yes Yes Yes Yes
Married  ×  num chil No No No Yes Yes Yes
All FEs interacted Yes Yes Yes Yes Yes Yes
Observations 152,056 64,017 64,017 24,933 23,615 23,615
Notes: Regressions are fixed-effect (or conditional) logistic regressions with coefficients reported as odds ratios.
Immigrants are excluded from all regressions. Robust standard errors are reported in brackets; all regressions are
clustered at the country level.
Table 4—WVS Countries with Large Within-Country FTR Differences
Strong FTR
Country Weak-FTR languages Percent Strong-FTR languages Percent Coef. SE N
Burkina Faso Dyula 16 French, Fula, Moore 84 0.687 [0.386] 137
Estonia Estonian 78 Russian 22 0.000 [0.000] 31
Ethiopia Amharic, Oromo, Sidamo 78 Chaha, Gamo, Tigrinya 22 0.837 [0.366] 208
Malaysia Malay, Mandarin 87 English, Tamil 13 0.745 [0.232] 449
Nigeria Yoruba 30 English, Hausa, Igbo 70 0.758 [0.354] 121
Singapore Malay, Mandarin 63 English, Tamil 37 0.813 [0.149] 664
Switzerland German 52 French, Italian 48 0.360 [0.133]*** 171
Notes: Coefficients (reported as odds ratios) are from fixed-effect (or conditional) logistic regressions with the
same specification as regression 6 in Table 3, but restricted to individual countries with significant within-country
variation in FTR strength. Listed languages are the most common weak- and strong-FTR languages in that country;
percents are the share of that country’s WVS sample that speak weak- and strong-FTR languages. Immigrants are
excluded from all regressions.
across individuals the belief that saving is an important value is almost completely
uncorrelated with the FTR of their language (corr = −0.07).
Parameter estimates from this first set of regressions indicate that a language’s
FTR is an important predictor of savings behavior. This effect is both large (larger
than that of other widely-studied variables) and survives an extremely aggressive
set of controls. Interestingly, this correlation is statistically independent of what was
Table 5—Additional Within-Country Control Regressions in the WVS
Saved
(1) (2) (3) (4) (5)
Strong FTR 0.857 0.676 0.669 0.537 0.539
[0.297] [0.096]*** [0.098]*** [0.111]*** [0.111]***
Unemployed 0.693 0.632 0.689 0.750 0.749
[0.046]*** [0.156] [0.044]*** [0.068]*** [0.067]***
Trust 1.072 1.277 1.084 1.066 1.067
[0.047] [0.139]** [0.045] [0.050] [0.050]
Family is important 0.963 0.792 0.953 0.991 0.990
[0.060] [0.128] [0.057] [0.069] [0.069]
Saving is important 1.124 0.978 1.110 1.056
(to teach children) [0.047]*** [0.080] [0.044]*** [0.060]
Language share 0.769 0.700 0.699
[0.120] [0.129] [0.129]
FTR share 1.016 0.475 0.469
[0.184] [0.190] [0.188]
Full set of FEs
from reg 5 in Table 3 Yes Yes Yes Yes Yes
Religion FEs No No No Yes Yes
All FEs interacted Yes Yes Yes Yes Yes
Country’s FTR variation < 5% (69) > 5% (7) All All All
Observations 21,834 1,781 23,615 13,245 13,245
Immigrants are excluded from all regressions. Robust standard errors are reported in brackets; all regressions are
designed to be a good marker of saving and thrift as a cultural value. This suggests
that the channel through which language affects the propensity to save is largely
independent of the saving as a self-reported value. Later, I will discuss what this
non-attenuation result suggests about the causal link between language and savings
behavior.
Next, I look at which countries in the WVS have numerous native speakers of
both weak- and strong-FTR languages. Figure 1 shows the percent of households
who reported savings for countries in the WVS, organized by what percent of the
country’s survey respondents report speaking a strong-FTR language at home.
As Figure 1 shows, the between-country relationship between savings and lan-
guage is both clear and highly significant in the WVS. However, the vast majority
of countries (69 of 76) have basically no intra-country variation in FTR strength.
This is because in most countries one language dominates, and in many multilingual
countries, those languages share a common FTR structure. For example, though
Canada has both English and French speaking populations, French and English are
both strong-FTR languages.
In 7 of 76 WVS countries however, both weak- and strong-FTR speakers are a sig-
nificant share of natives. These seven countries provide the majority of identification
for my within-country regressions. Table 4 enumerates these countries, and reports
the coefficient on Strong FTR when regression 6 from Table 3 is estimated in that
60
Switzerland
50
Percent of households saving
Malaysia
Singapore
40
Estonia Nigeria
30
Burkina Faso
10 countries have almost
no strong-FTR speakers
20
Coef: –0.1558
T = –4.19 59 countries have almost
Ethiopia P < 0.001 no weak-FTR speakers
10
0
0 10 20 30 40 50 60 70 80 90 100
Share of households speaking a strong-FTR language (percent)
Figure 1. Between-Country Savings and Language in the WVS
Notes: Figure 1 plots the least-squares regression of the percent of a country which reports saving on the percent of
that country which speaks a strong-FTR language at home. The large number of countries with extreme strong-FTR
percentages (< 5 percent and > 95 percent), are summarized by their means and standard errors.
country. I also report the percents of each country’s sample that speak strong- and
weak-FTR languages, the most common languages they speak, and the sample size
of each country-specific regression.
Remember that these coefficients represent the odds ratio of savings, for strong-
over weak-FTR families. Though small samples reduce statistical significance,
consistent with my overall effect, all 7 regressions display coefficients less than 1,
(strong-FTR families save less often than their weak-FTR counterparts). The coeffi-
cient in Estonia is 0 because in the sample of matched Estonian families, no Russian
speakers reported saving. Other than this outlier, (which is driven by that regres-
sion’s small sample size) the estimated effect is remarkably stable across this set of
countries, which span multiple continents, regions, and sets of languages.
To confirm this and to explore the robustness of my initial results to additional
controls, I estimate an additional set of regressions summarized in Table 5. First, I
estimate the regression with a full set of controls (regression 6 in Table 3) separately
in the 69 countries with little, and the 7 countries with sizable within-country FTR
variation (columns 1 and 2 in Table 5, respectively). I also examine whether these
results are explained by the share of a country speaking a language or languages of a
particular FTR level. Finally, I add fixed-effects for self-reported religious denomi-
nation (74 in total), interacted with all of my previous fixed effects.
Regressions 1 and 2 confirm that the majority of the identification for my within-
country regressions come from the seven countries enumerated in Table 4. The coef-
ficient of 0.676 in regression 2 is statistically indistinguishable from the coefficient
of 0.691 I measure when that regression is run on the whole sample.
Returning to the whole sample: as an additional control, regressions 3, 4, and 5
add controls for the share of a country that speaks a household’s language, and what
share speak a language with the same FTR level. Neither of these attenuate the effect
Table 6—Household Retirement Assets (SHARE)
IHS( _ DI )
  RA    
(1) (2) (3) (4) (5)

Strong FTR −0.390 −0.370 −0.444 −0.386 −0.356
[0.017]*** [0.024]*** [0.028]*** [0.047]*** [0.075]***
Fixed effects:
Age Yes Yes Yes Yes Yes
Country × wave Yes Yes Yes Yes Wave
Income No Yes Yes Yes Yes
Education No No Yes Yes Yes
Married × num chil No No No Yes Yes
All FEs interacted Yes Yes Yes Yes Yes
Countries All All All All BE and CH
Observations 39,665 39,665 39,665 39,350 5,937

F stat 529.55 234.93 255.48 68.90 23.95
Notes: Regressions are fixed-effect OLS regressions where the dependent variable is the inverse-hyperbolic sine of
net household retirement assets divided by average national disposable income. Immigrant households are excluded
from all regressions. Robust standard errors are reported in brackets; all regressions are clustered at the country
level except regression 5, which is clustered at the household level.
of language. Regressions 4 and 5 include fixed effects for religious denomination

(74 in total), interacted with all of my previous fixed effects. While the addition of
religion significantly reduces the usable sample, its inclusion does not attenuate the
effect of language; comparing regression 3 to 4, the measured effect actually grows
by 11 percent. Comparing regression 4 to 5 replicates my earlier non-attenuation
finding: the addition saving as a self-reported value does not affect the effect of
language.
B. Language and Retirement Assets in Europe
If individuals who speak strong-FTR languages save less in any given year, then
we would expect them to accumulate less savings over time. My next set of regres-
sions examines the cumulative retirement assets of individuals in the retired house-
holds in the SHARE.
Table 6 summarizes regressions that estimate the equation
(  )
rait
(11) IHS _
        = α + β1  StrongFTR + β2(F  ex
it   × 
F  en
it   × 
F  ct  ) + εit,
incct
where
IHS(x) = log x + (x  2 + 1)2   . (  _ 1

)
     
In equation (11) the dependant variable is IHS( _

ct )
ra
  incit    . The numerator rait is the
estimated value of a retired household’s net worth, including all real assets (homes,
businesses, and cars), and financial assets (money, stocks, bonds, and life i nsurance),
minus any debt. In order to make this comparable across families living in different
countries, I divide raitby incct, that country’s average disposable income.30 Finally,
in order to assure that these regressions are not being driven by outliers, I apply
IHS(x), an inverse-hyperbolic-sine transformation. Except for small values of x,
this transformation approximately equals log(2x), or log(2) + log(x), and so coef-
ficients can be interpreted in exactly the same way as with a standard log transfor-
mation. Unlike the log transformation though, IHS(x) is defined for both zero and
negative values of x, which are common in wealth data.31
Unfortunately, unlike the WVS, the SHARE does not ask households what lan-
guage they speak at home. Therefore, the main variable of interest Strong FTR is
coded using the language that the head of household asked to take the survey in (the
SHARE attempts to offer the survey in any of a country’s national languages).
Mirroring my earlier regressions in the WVS, the F variables are sets of fixed
effects that are jointly interacted to form groups similar to those in my analysis of
the WVS. That is, households are compared only with others who are identical on
every F variable, but who asked to take the survey in a different language. Empirical
estimates of equation (11) are presented in Table 6; all coefficients can be inter-
preted as percent changes in retirement savings.32
Regressions 1 through 5 show my predicted effect; retired households that speak
strong-FTR languages have around 39 percent less by the time they retire. These
regressions are identified by the fact that Belgium has large Flemish (weak-FTR)
and French (strong-FTR) speaking populations, and Switzerland has large German
(weak-FTR), and French, Italian, and Romansh (strong-FTR) speaking populations.
Comparing regressions 4 and 5, we see that the differences overall effect appear to
be roughly the same size as the differences between different FTR groups within
Belgium and Switzerland.
Table 7 summarizes regressions that increase the level of spatial control by includ-
ing fixed effects for intra-country regions. This allows us to examine whether lan-
guage may be proxying (even within country) for unobserved differences between
regions, counties or even cities. If for example, families tend to segregate across
regions by language, then I may be attributing institutional differences between
regions to language.
Comparing regressions 1 and 2 (in Belgium) and regressions 3 and 4 (in
Switzerland) shows that the addition of finer spatial controls (in the form of 11
Belgian and 7 Swiss region dummies) do not appear to attenuate the effect of lan-
guage on retirement savings. For example, Brussels has both a large Dutch speaking
(weak-FTR) and a large French speaking (strong-FTR) population. Regressions 5
and 6 show that even when comparing demographically similar families living in
Brussels, language appears to have a strong effect on retirement savings. Together,
30
Average disposable income is per head, with PPP and in 2005 Euros, as collected by the OECD.
31
I am indebted to Frances Woolley, who suggested I use the inverse-hyperbolic-sine transformation, and pointed
me to her work on the transformation and its use with wealth data. See Woolley (2011) and Burbidge, Magee, and
Robb (1988) for a discussion of the relative advantages of various transformations. All of the regressions I report
using this transformation provide qualitatively and statistically similar results when run without any transformation,
either as a ratio to disposable income or simply in levels.
32
Details on variable construction: Age is coded in ten-year bins, Income is coded as an intra-country decile,
and Education falls within one of eight categories provided in the SHARE. For more details on the construction of
variables and the measuring of household net-worth in the SHARE, see Börsch-Supan and Jürges (2005).
Table 7—Household Retirement Assets (SHARE, BE, and CH)
IHS( _ DI )
  RA    
(1) (2) (3) (4) (5) (6)

Strong FTR −0.459 −1.637 −0.374 −0.440 −1.621 −1.571
[0.052]*** [0.515]*** [0.132]*** [0.194]** [0.457]*** [2.749]
Fixed effects:
Age Yes Yes Yes Yes Yes Yes
Wave Yes Yes Yes Yes Yes Yes
Income Yes Yes Yes Yes No Yes
Education Yes Yes Yes Yes No Yes
FEs interacted Yes Yes Yes Yes Yes Yes
Sub-reg FEs 1 11 1 7 1 1
Sample BE BE CH CH Brussels Brussels
Observations 4,410 4,409 1,553 1,553 148 148

F stat 78.53 11.36 7.99 1.99 12.56 0.33
Notes: Regressions are fixed-effect OLS regressions where the dependent variable is net household retirement
assets divided by average national disposable income, with a inverse-hyperbolic-sine transformation. Immigrant
households are excluded from all regressions. Robust standard errors are reported in brackets; all regressions are
clustered at the household level.
these regressions suggest that the language effects I find are not due to spatial differ-
ences, at least not at the level that can be measured in the SHARE.
C. Language and Health in the SHARE
In addition to measuring household wealth, the SHARE also asks about health
behaviors and records several measures of physical health. If languages affect their
speakers’ intertemporal beliefs, this would also affect health behavior and long-run
health. More specifically, if obligatory FTR reduces the psychological importance
of the future, we would predict that it would lead to more smoking, less exercise,
and worse long-run health.
To investigate this, Table 8 summarizes regressions investigating the effect of
FTR on health variables found in the SHARE. Some of these measures are binary,
such as ever having smoked heavily, remaining physically active, and being medi-
cally obese. For these regressions I estimate fixed-effect logit model similar to equa-
tion (10). The other measures I examine, walking speed, grip strength, and peak
expiratory flow, are commonly studied measures of long-run health. These measure
the speed at which a person comfortably walks, the maximum amount of force they
can apply while squeezing a dynometer, and their maximum exhalatory air flow
(lung strength). For these regressions I estimate fixed-effect OLS regressions.
Regression 1 indicates that a strongly grammaticalized FTR leads to a 24 percent
higher probability of having ever smoked (daily for a year or more). This is con-
sistent with my findings on savings if the decision to smoke trades off immediate
benefits versus future health costs. Similarly, regression 2 indicates that a strong-
FTR language leads to a 29 percent lower probability of being physically active.
Regressions 3, 4, 5, and 6 examine the effect of strong-FTR on long-run measures
Table 8—Health Behaviors and Measures of Health (SHARE)
Smoked Phys. act. Obesity Walk sp. Grip str. Peak flow
(1) (2) (3) (4) (5) (6)
Strong FTR 1.241 0.709 1.131 −0.028 −0.899 −16.083
[0.042]*** [0.025]*** [0.007]*** [0.101] [0.049]*** [2.806]***
Full set of FEs
from reg 4 Table 6 Yes Yes Yes Yes Yes Yes
Observations 15,750 9,135 11,958 6,038 51,571 26,836

R2 0.85 0.84 0.73
Notes: Regressions 1, 2, and 3 are fixed-effect (or conditional) logistic regressions with coefficients reported as
odds ratios. The dependent variables are: having smoked daily for a year or more, engaging in regular physical
activity, and medically obese. Regressions 4, 5, and 6 are fixed-effect OLS regressions for measures of old-age
health; walking speed (m/sec), grip strength (kg), and peak expiratory flow(L/min). Immigrants are excluded from
all regressions. Robust standard errors are reported in brackets; all regressions are clustered at the country level.
of health. While there appears to be no effect on walking speed, speaking a strong-

FTR language is associated with a 13 percent higher probability of being medically
obese, a reduction in grip strength of almost a kilogram, and a reduction in peak
expiratory flow of 16 liters per minute.
D. Language and Health in the DHS
The detailed health measures found in the SHARE allow me to investigate both
health behaviors, and long-run markers of health. These comparisons however, are
limited to retired citizens of developed Western-European countries, and in the pres-
ence of country fixed effects are identified almost entirely by retired households
from Belgium and Switzerland.
To investigate whether these results generalize, I run similar health regressions
in data from the MEASURE DHS project. Of the countries covered by the DHS,
four countries contain significant (>5 percent) native within-country FTR variation.
These countries are Burkina Faso, the Democratic Republic of the Congo, Ethiopia,
and Nigeria.33 I use the DHS surveys for these countries in logistic health regres-
sions similar to those run in the SHARE. Table 9 reports the results of these regres-
sions on smoking, obesity, contraception, and condom use. If speaking a strong-FTR
language reduces a person’s concern for the future, we should expect it to increase
smoking and obesity, and decrease family planning and safe-sex behavior.
Regressions 1 through 6 are consistent with my general hypothesis, and quantita-
tively similar to the results I obtain in the SHARE. Speaking a strong-FTR language
is associated with a 20 percent greater likelihood of smoking, and a 17 percent
33
In order to match as closely as possible the health regressions I run in the SHARE, I use the DHS survey that
falls between 2004 and 2006 for each country, with two changes in specification. Age in the DHS data comes coded
in five-year bins, and I maintain that greater level of specificity than I used in the SHARE (ten-year bins). Also, I
include fixed effects for self-reported religion, which was not measured in the SHARE (but is in the WVS, and was
included in those regressions).
Table 9—Health Behaviors in Developing Countries (DHS)
Smokes Obesity Cntcpt. Cntcpt. Condom Condom

(1) (2) (3) (4) (5) (6)
Strong FTR 1.199 1.169 0.533 0.575 0.810 0.838
[0.104]** [0.058]*** [0.025]*** [0.029]*** [0.052]*** [0.069]**
FTR × sex 0.726 0.926
[0.088]*** [0.122]
Full set of FEs
from reg 4 Table 6 Yes Yes Yes Yes Yes Yes
Religion FEs Yes Yes Yes Yes Yes Yes
Observations 13,252 27,706 32,064 32,064 11,201 11,201
The dependent variables are currently smokes, being medically obese, uses any form of contraception, and used a
condom during last intercourse. Robust standard errors are reported in brackets.
greater likelihood of being obese. In regressions 3 through 6 I examine sexual health

behaviors that have a strong future-component, reporting using contraception for
family planning, and reporting having used a condom during your last sexual inter-
course. Speaking a strong-FTR language significantly depresses both behaviors.
Also, this effect appears stronger for men than for women, consistent with a large
literature in development that finds men may have greater decision-making author-
ity over these behaviors.
E. Linguistic Effects on National Savings Rates in the OECD
The evidence on both individual and household behavior we have presented so

far supports the hypotheses that strong-FTR languages induce less future-oriented
choices by its speakers. If my hypothesis about language and willingness to save
is true however, it would also have implications for aggregate behavior. It seems
natural to expect that countries in which strong-FTR languages are spoken would
have both lower equilibrium household savings, and (to the degree governments
aggregate individual preferences) government savings.34 Figure 2, which graphs the
relationship between language and savings rates for OECD countries (without any
controls), suggests that the results we find among households also seem to hold for
national savings rates.
To see whether this trend survives a basic set of controls, Table 10 summarizes
a first set of regressions that comprise a more careful test of this prediction. These
34
This prediction does not immediately follow from theory, however. Samuelson (1937) showed that when the
duration of a potential project is fixed, the value of that project may not be even weakly decreasing in the interest
rate. Arrow and Levhari (1969) established that if an agent controls when a project terminates, then in deterministic
settings the natural monotonic relationship must hold; the value of investment in projects must be monotonically
decreasing in the interest rate. In Hicks’ (1973) book Capital and Time, this is referred to as the Fundamental
Theorem of Capital. Under the conditions for which this relationship holds then, it is natural to predict that countries
with strong-FTR languages will, on average, save less.
45
On average, countries which speak strong-FTR
Average total savings rate (percent GDP)
40 languages save 4.75% less.

t = 2.77, p = 0.009
35
Light-bar countries speak
weak-FTR languages
30
Dark-bar countries speak
strong-FTR languages
25
20
15
10
H pa c
rla ile
Tu ly
un in
te gd l
ep n
Fr key
er nd
Au ve ia
el e
or and
e o
Po and
Sl ust d
e
ep ia
itz ela y
en um
G eal a
R ton n
U Ki Isra l
s nia
ew a ry
D lgi ia
M ark
Ita y
ch Ja nd
nl s
Fe K urg
r a
m d
re s
et Cblic
St m
or n
ni n e
ga
S li
Sw Ir wa
an
Fi nd
G ate
R pa
Ic nc
Sw xic
A an
ec
Z nad
ak Es de
de ore
er an
N atio
o r
Be tral
N C ga
d o
ub
he h
la
tu
m
an th o
r
a
l
e
si ou b
us S xem
d
Lu
te
ov
ze
ni
Sl
C
U
R
Figure 2. OECD Savings Rates and Language, 1985–2010
Notes: Figure 2 shows average total savings rates, accounting for both private and government consumption. Both
Switzerland and Belgium have significant within-country FTR variation; for simplicity they are shaded according
to their majority-FTR status. Difference in means are computed using a OLS regression where observations are
regressions closely follow Barro and McDonald (1979), who run similar regressions
on the same OECD national savings data that we investigate here. The basic func-
tional form of these regressions is
(12) _
  ( 
Y − C
Y
   
it
)
   = α0 + α1(StrongFTR)i + α2 _
1
Y it (  )
        + α3 _(  )
t−1
Y
       
Yt it
+ α4(CAGR)i + εit,
where annual observations for each country in the OECD are indexed by coun-
try i = 1, … , 35 and year t = 1970, … , 2009.35 Most importantly: C is total con-
sumption (including government expenditure) while Y is GDP, CAGR is the average
growth rate of the country from 1993 to 2009 (the earliest date for which data is
available for all countries), and StrongFTR is weighted by the percent of the coun-
try’s population reports speaking each of their major languages.36
This form of this savings equation is a simple linear relation motivated by the
Life-Cycle Hypothesis (LCH) of savings (see Modigliani 1986 for a review of the
LCH). Notice that as equation (12) is written, all terms in the savings equation
except (1/Y )itimply that a savings function that is homogeneous of degree 0, which
35
Details on the construction of each variable can be found in Appendix A.
36
Multilingual countries’ FTR values need to be weighted by the shares of their populations speaking each
language. I used the Ethnologue database to do this, see Lewis (2009) for details.
Table 10—Gross Domestic Savings Rates in the OECD
GDSRt
(1) (2) (3) (4) (5) (6)
Strong FTR −5.272 −5.212 −5.245 −6.397 −5.821 −5.924
[1.798]*** [1.769]*** [1.948]** [1.774]*** [2.378]** [2.457]**
PCGDPt−1 /PCGDPt −32.504 −35.035 −45.985 −41.852 −41.285 −32.396
[7.908]*** [8.930]*** [14.742]*** [13.215]*** [12.065]*** [10.819]***
CAGR −0.045 −0.009 0.141 0.154 0.169 0.159
[0.128] [0.169] [0.284] [0.265] [0.255] [0.228]
Unemploymentt(percent) −0.411 −0.367 −0.226 −0.190 −0.170 −0.171
[0.141]*** [0.133]*** [0.142] [0.130] [0.125] [0.118]
Oldt (percent) −1.342 −1.382 −1.386 −1.240 −1.100 −1.165
[0.284]*** [0.298]*** [0.318]*** [0.319]*** [0.283]*** [0.299]***
Youngt(percent) −0.711 −0.663 −0.618 −0.367 −0.205 −0.155
[0.162]*** [0.184]*** [0.239]** [0.212] [0.302] [0.351]
1/PCGDPt −28.119 −66.266 −109.761 −130.260 −133.831
[49.798] [65.211] [57.808] [71.601] [68.227]
Soc Sect(percent GDP/old) −1.942 −1.909 −1.622 −5.457
[2.252] [2.040] [2.153] [3.395]
Protestant −4.102 −4.229 −5.501
[1.427]*** [1.850]** [1.196]***
French legal origin −0.424 0.358
[2.425] [3.401]
German legal origin 1.542 2.197
[3.316] [4.283]
Scandanavian legal origin 0.395 0.503
[3.508] [3.343]
Distance from equator 1.371
[3.106]
Continent FEs: No No No No No Yes
Observations 904 904 614 614 614 614

R2 0.40 0.40 0.43 0.51 0.52 0.59
Notes: Regressions are OLS regressions where the dependent variable is a country’s gross domestic savings rate
in year t. Observations are for OECD countries from 1970 to 2009. Protestant is a binary variable that measures if
the country is majority protestant or not. Robust standard errors are reported in brackets and clustered at the coun-
try level.
is to say that the savings rate is independent of the level or unit of income. This
assumption has theoretical support in the LCH model, and allows for a specification
in which units of measurement do not need to be comparable across countries. It
may be violated if, as Feldstein (1980) points out, higher incomes lead to a increase
in the share of life spent in retirement. This leads to the presence of the 1/Yitterm,
which can test for such effects as measured by a positive α 2. Essentially this term
allows the marginal propensity to consume out of income to differ by the level of
development of a country. In addition, OECD data allows for the inclusion of a num-
ber of important demographic controls:
α5(Unemployment)it + α6(Old)it + α7(Young)it + α8(SocSec)it .
These control for the unemployment rate, the fraction of the population that are over
65, the fraction under 15, and the per capita fraction of GDP spent on social security
payments (defined as percent GDP spent on disability, old age, and survivors ben-
efits divided by the fraction of the population that are over 65).
Another possible concern with cross-country regressions may be that the FTR
strength of countries is spatially correlated. In Western Europe for example, most
strong-FTR countries are in the northern half of the continent. This leads to the
possibility that (at least in Western Europe), the effects I attribute to strong-FTR
could actually be due to correlated spatial factors (like climate or distance from
Mediterranean trade routes) Similar stories might also invalidate my results on other
continents. The inclusion of continent fixed effects and a country’s Latitude allow
me to investigate whether effects like these are biasing my results.
Empirical estimates of equation (12) are presented in Table 10.
Regression 1 estimates a version of equation (12) that is fully homogeneous of
degree 0, while regressions 2 through 6 add 1/PCGDPt , which allows savings rates
to vary with the size of the economy. These regressions suggest that countries with
a strong-FTR language save on average around five percentage points less per year
than do countries with a weak-FTR language, a result consistent with my earlier
results on household savings and health measures. Regressions 5 and 6 add con-
trols commonly found in the literature on economic growth: Protestantism and a
country’s legal origin.37 Regression 6 adds two measures to help control for spatial
correlation, continent fixed effects, and the distance from a country’s capital to the
equator in thousands of miles. Overall, the measured effect of FTR on national sav-
ings rates is stable to the inclusion of these controls. Weak-FTR countries appear
to save on average 6 percent more of their GDP per year than their strong-FTR
counterparts.
F. Language and Savings in the OECD: Robustness Checks
To get a sense of the stability of my measured effect over time, I re-estimate

equation  (12) separately for each decade that OECD data is available. These esti-
mates are reported in Table 11.
While statistical power becomes an issue when subdividing these data, the effect
of language on savings appears stable across time, and is significant in every decade.
Earlier regressions have fewer observations due to OECD expansion in the 1980s
and early 1990s. Increasing membership in the OECD also makes it hard to compare
coefficients across time periods; however in a pooled regression the interactions
between language and decade dummies are insignificant.
These OECD regressions have the benefit of comparing relatively similar coun-
tries: the member nations of the OECD are developed economies dedicated to open
markets and free trade. By merging World Bank national savings data with the WVS
however, we can investigate whether these cross-country national savings results
extend to less developed countries. In addition, this allows us to examine whether
37
A large literature has argued that common-law countries provide stronger protection of outside investors from
expropriation by corporate insiders, and that this and other features of a legal system are largely determined by a
country’s legal origin; see La Porta, Lopez-de-Silanes, and Shleifer (2008) for an excellent survey.
Table 11—Gross Domestic Savings Rates in the OECD by Decade
GDSRt
(1) (2) (3) (4) (5)
Strong FTR −5.212 −4.194 −5.458 −5.142 −6.975
[1.769]*** [1.750]** [1.805]*** [2.043]** [2.667]**
PCGDPt−1/PCGDPt −35.035 −46.258 −64.147 −49.460 −18.060
[8.930]*** [12.536]*** [15.494]*** [19.760]** [8.391]**
CAGR −0.009 0.397 0.313 0.121 −0.056
[0.169] [0.571] [0.329] [0.371] [0.158]
Unemploymentt(percent) −0.367 −0.544 −0.360 −0.229 −0.321
[0.133]*** [0.307] [0.189] [0.192] [0.153]**
Oldt (percent) −1.382 −1.298 −1.178 −1.803 −2.019
[0.298]*** [0.552]** [0.321]*** [0.382]*** [0.513]***
Youngt(percent) −0.663 −0.624 −0.213 −0.810 −1.237
[0.184]*** [0.464] [0.320] [0.257]*** [0.230]***
1/PCGDPt −28.119 112.425 −23.892 −49.676 −66.561
[49.798] [78.480] [29.367] [53.792] [85.171]
Years: All 1970–1979 1980–1989 1990–1999 2000–2009
Observations 904 103 185 290 326

R2 0.40 0.69 0.64 0.45 0.39
Notes: Regressions are OLS regressions where the dependent variable is a country’s gross domestic savings rates in
year t. Robust standard errors are reported in brackets and clustered at the country level.
these cross-country regressions survive the types of culture and values controls that
are present in the WVS. Table 12 reports the results of these regressions.
The results of these regressions suggest that the national savings results I find
are not limited to the developed nations of the OECD, and also survive the cultural
controls that are found in the WVS. Overall, if anything the effect of language on
national savings appears stronger (in percentage points) among developing than
developed nations, and are not significantly attenuated by the inclusion of culture
and values controls.
V. Discussion
A. Language, Thought, and Behavior
The idea that language can impact the way people think and act has a rich history
in linguistics, philosophy, and psychology. Saussure, the founder of both structural
linguistics and semiotics, characterized reality as an unstructured phenomena that
is discretized and organized by language, writing: “if words stood for pre-existing
entities they would all have exact equivalents in meaning from one language to the
next, but this is not true” (Saussure 1916). In his Tractatus Logico-Philosophicus,
Wittgenstein (1922) formulates a theory of language as the means by which people
both picture and reason about reality, famously concluding: “Wovon man nich spre-
chen kann, darüber muss man schweigen” (Whereof one cannot speak, thereof one
Table 12—Gross Domestic Savings Rates in the WVS
GDSRt
(1) (2) (3) (4) (5) (6)
Strong FTR −13.848 −15.545 −12.253 −11.328 −8.442 −11.002
[3.303]*** [4.814]*** [3.337]*** [3.320]*** [2.980]*** [3.611]***
PCGDPt−1/PCGDPt 47.570 19.905 15.108 15.616 13.488 11.032
[35.330] [28.120] [25.197] [23.366] [21.735] [19.177]
Oldt (percent) −2.363 −1.718 −1.916 −2.112 −1.144 −1.064
[0.685]*** [0.839]** [0.730]** [0.687]*** [0.774] [1.312]
Youngt(percent) −1.274 −0.736 −0.813 −0.891 −0.617 −0.523
[0.409]*** [0.498] [0.501] [0.512] [0.475] [0.693]
French legal origin −7.676 −3.302 −7.578 −4.191 −2.189
[2.843]*** [2.828] [4.887] [4.831] [6.080]
German legal origin −9.937 −6.735 −11.716 −6.883 −6.450
[6.790] [4.980] [4.828]** [5.204] [8.180]
Scandanavian legal origin −7.430 −3.196 −6.432 1.139 omitted
[7.248] [5.326] [5.355] [5.580]
1/PCGDPt −4.455 −4.819 −5.102 −5.477 −4.811
[1.726]** [1.781]*** [1.766]*** [1.954]*** [2.318]**
Unemploymentt(percent) −0.724 −0.587 −0.455 −0.252
[0.193]*** [0.225]** [0.207]** [0.250]
Real interest ratet −0.199 −0.219 −0.217 −0.213
[0.108] [0.092]** [0.076]*** [0.081]**
Legal rights index −0.899 −0.084 0.236
[0.999] [1.059] [1.280]
Trustt 2.947 −0.787 −5.474
[9.244] [8.738] [11.585]
Family is importantt 47.163 42.008 49.974
[15.877]*** [13.648]*** [17.618]***
Continent FEs: No No No No Yes Yes

PCGDP All All All All All < 5, 000
Observations 120 120 113 113 113 73

R2 0.20 0.32 0.51 0.55 0.62 0.65
Notes: Regressions are OLS regressions where the dependent variable is a country’s gross domestic savings rates in
year t. Observations are for the countries in the WVS countries over three waves, from 1994 to 2008. Robust stan-
dard errors are reported in brackets and clustered at the country level.
must be silent). The idea that language can influence thought has become know as
the Sapir-Whorf hypothesis (SWH, Whorf 1956), and has generated several interest-
ing lines of research in linguistics and psychology.38 My hypothesis can be thought
of as an instance of the SWH, and is to my knowledge, the first to connect language
structure and decision making.
Skepticism of the Weak Sapir-Whorf Hypothesis.—While many studies support at

least a weak form of the SWH, there are a number of scholars who argue that on bal-
ance, the idea that cognition is shaped by language is misguided. Most prominently,
38
See Scholz, Pelletier, and Pullum 2011 for a review of Sapir-Whorf hypotheses.
in his seminal work Syntactic Structures, Chomsky (1957) argues that humans have
an innate set of mechanisms for learning language, and that this constrains all human
languages to conform with a “universal grammar.”
Taken in strong form, a universal grammar would largely eliminate the scope for
language to affect cognition. In The Language Instinct, Pinker (1994) argues exactly
this: that humans do not think in the language we speak in, but rather in an innate
“mentalese” which precedes natural language. He concludes that: “there is no scien-
tific evidence that languages dramatically shape their speakers’ ways of thinking”
(emphasis mine).While a rich literature since 1994 has disputed this claim, support
for the SWH remains an hotly debated topic.
Language Acquisition and Future-Time Reference.—Important for evaluating my

hypothesis, several studies have looked at differences between children learning
weak- and strong-FTR languages. Harner (1981) finds that among English-speaking
children, the use of the future tense begins by age three and is relatively developed
by age five. Szagun (1978) finds that the time-path of this development is identi-
cal in matched pairs of English and German children, with these pairs of children
showing no discernible difference in the rate at which they acquire and use FTR.
Differences between English (strong-FTR) and German (weak-FTR) were reflected
in Szagun’s study, but only among adults: the German-speaking parents of the chil-
dren Szagun studied used FTR much less often than their English-speaking coun-
terparts. These similar development paths suggest that the differences that I find
between weak- and strong-FTR language speakers do not reflect innate cognitive
nor early cultural differences between speakers of different languages, at least as
reflected in the development of children through age five.
Work on Language in Economics.—Work on language in economics has primar-

ily focused on whether language, either by evolution or design, maximizes some
objective function. The earliest example of this is Marschak (1965) which asks both
which traits will be selected as languages evolve, and what objectives policy mak-
ers should have in mind when shaping a language, either directly (as in the case of
the Académie française)39 or through educational policy. Closest in objective to this
paper, Rubinstein (2000) studies a model in which decision makers use language to
both perceive and verbalize decisions. It follows that: “interesting restrictions on the
richness of a language can yield interesting restrictions on the set of an economic
decision maker’s admissible preferences.” This “expressibility effect” is essentially
a much stronger form of what I test for here: the ability of language to affect beliefs
and behavior.
Work on Development and Growth.—There is also a broad and ongoing debate

as to why similarly-situated nations and societies can differ so greatly in their eco-
nomic development and wealth. Jared Diamond is probably best associated with
the geographer/biologist’s view that these differences are mainly due to geography,
39
The Académie francaise is made up of 40 members (immortels) who are elected to life terms. The Académie
is France’s official authority on the vocabulary and grammar of the French language, and publishes the Dictionnaire
de l’Académie française, the official dictionary of the French language.
climate, and the ecology of animal domestication (Diamond 2005). Historian David
Landes argues that deep seated cultural factors affect the ability of societies to exploit
science, technology, and markets (Landes 1998). Finally, social scientists such as
Acemoglu, Robinson, and Shleifer have argued for the central role of institutions
in providing the right incentives for innovation and good government (Acemoglu
and Robinson 2012; La Porta, Lopez-de-Silanes, and Shleifer 2008). All of these
theories were developed by comparing similarly situated nations and societies that
have experienced divergent economic outcomes. While my findings are not a theory
of development, they do suggest that language structure is an important factor to
account for when making such comparisons.
VI. Conclusion
Overall, my findings are largely consistent with the hypothesis that languages with
obligatory future-time reference lead their speakers to engage in less future-oriented
behavior. On savings, the evidence is consistent on multiple levels: at an individual’s
propensity to save, to long-run effects on retirement wealth, and in national savings
rates. These findings extend to health behaviors ranging from smoking to condom
use, as well as to measures of long-run health. All of these results survive after
comparing only individuals who are identical in numerous ways and were born and
raised in the same country.
One important issue in interpreting these results is the possibility that language
is not causing but rather reflecting deeper differences that drive savings behavior.
These available data provide preliminary evidence that much of the measured effects
I find are causal, for several reasons that I have outlined in the paper. Mainly, self-
reported measures of savings as a cultural value appear to drive savings behavior,
yet are completely uncorrelated with the effect of language on savings. That is to
say, while both language and cultural values appear to drive savings behavior, these
measured effects do not appear to interact with each other in a way you would
expect if they were both markers of some common causal factor.
In addition, differences in the use of FTR do not seem to correspond to cognitive
or developmental differences in the acquisition of language. This suggests that the
effect of language that I measure occurs through a channel that is independent of
either cultural or cognitive differences between linguistic groups.
Nevertheless, the possibility that language acts only as a powerful marker of some
deeper driver of intertemporal preferences cannot be completely ruled out. This pos-
sibility is intriguing in itself, as the variation in future-time reference that identifies
my regressions is very old. In Europe for example, most Germanic and Finno-Ugric
languages have been futureless for hundreds of years. Indeed, Dahl (2000) suggests
that proto-Germanic was futureless at least 2,000 years ago.
Appendix A: Data
A. Data Statements
This paper uses data from SHARE release 2.3.1, as of July 29, 2010. SHARE data
collection in 2004–2007 was primarily funded by the European Commission through
its fifth and sixth framework programes (project numbers QLK6-CT-2001-00360;

RII-CT-2006-062193; CIT5-CT-2005-028857). Additional funding by the US
National Institute on Aging (grant numbers U01 AG09740-13S2; P01 AG005842;
P01 AG08291; P30 AG12815; Y1-AG-4553-01; OGHA 04-064; R21 AG025169)
as well as by various national sources is gratefully acknowledged (see http://www.
share-project.org for a full list of funding institutions).
B. Wording of Questions in the WVS (Household Level Variables)
FAMSAVED: During the past year, did your family (read out and code one
answer):
1: Save money (23 percent)
2: Just get by (51 percent)
3: Spent some savings and borrowed money (14 percent)
4: Spent savings and borrowed money (12 percent)
For the regressions in this paper, this variable is coded as 1 if the family reported
saving money, and 0 otherwise.
TRUST: Generally speaking, would you say that most people can be trusted or
that you need to be very careful in dealing with people? (Code one answer):
1: Most people can be trusted. (26 percent)
2: Need to be very careful. (74 percent)
FAMILY: indicate how important it is in your life. Would you say it is

1: “Very important” (91 percent)
2: “Rather important” (8 percent)
3: “Not very important” (1 percent)
4: “Not at all important” (0.2 percent)
CHILDSAVE: Here is a list of qualities that children can be encouraged to learn

at home. Which, if any, do you consider to be especially important? Please choose
up to five! (Code five mentions at the maximum):
Independence Hard work

Feeling of responsibility Imagination
Tolerance and respect for other people Thrift, saving money and things (37 percent)
Determination, perseverance Religious faith
Unselfishness Obedience
UNEMPLOYED: Respondents are asked to chose one of three options:

1: Employed (52 percent)
2: Not in Labor Force (38 percent)
3: Unemployed (10 percent)
C. World Bank, World Development Indicators (WDI)
LEGAL ORIGIN VARIABLES: Following La Porta et al. 1997, many

cross-country analyses have included fixed effects for the origin of a country’s legal
Table A1—Data Summary: WVS, Country Level
Mean SD
WDI variables
GDSRt 18.2% 14.2%
PCGDPt(US$) $7,375 $10,080
PCGDPt−1/PCGDPt 0.967 0.043
Real interest ratet 6.28% 15.9%
Oldt 9.14% 4.87%
Youngt 26.9% 9.26%
Legal rights index (1 to 10) 5.99 2.31
WVS variables
Trust (country average) 25.9 14.9
Family (country average) 1.11 0.074
Legal origins (number of countries)

UK (19), French (37), German (16), Scandinavian (3)
Table A2—Data Summary: WVS, Household Level
Percent of households
Saved this year 23.0
Strong FTR (language spoken at home) 85.0
Sex (male) 48.2
Unemployed 9.84
Most people can be trusted (agree) 26.2
Mean SD
Age 40.5 16.1
Age finished school 19.3 6.35
Number of children 1.95 1.85
Family importance (1 “very”– 4 “not at all”) 1.10 0.351
Language share 76.1% 31.9%
FTR share 97.4% 11.2%
system. These controls account for the fact that legal rules protecting investors
appear to vary systematically among legal traditions, with La Porta et al. (1997)
arguing that the laws of common law countries (originating in English law) are
more protective of outside investors than civil law countries (originating in Roman
law). They argue that these protections limit the extent of expropriation of outside
investors by corporate insiders, and thereby promotes financial development. In this
paper I include fixed effects for French, German, and Scandinavian legal origins,
with English legal origins as the excluded category. See La Porta, Lopez-de-Silanes,
and Shleifer (2008) for a summary of this measure and the literature surrounding it.
LEGAL RIGHTS INDEX: The World Bank strength of legal rights index mea-
sures the degree to which a country’s collateral and bankruptcy laws protect the
rights of borrowers and lenders and thus facilitate lending. The index ranges
from 0 to 10, with higher scores indicating that these laws are better designed to
expand access to credit. For details, see the World Bank Doing Business Project,
(www.doingbusiness.org).
Table A3—Data Summary: SHARE, Household Level
Smoked 47.6
Physically inactive 89.4
Obesity (BMI > 30) 17.6
Mean SD
Household net worth (€) 333,417 € 1,183,231 €
Disposable income (€) 21,354 € 3,935 €
Age (years) 65.9 10.5
Education (years) 10.7 4.42
Walking speed (m/s) 0.692 0.373
Grip strength (kg) 34.5 12.1
Peak expiratory flow (L/min) 337 160
PCGDP is gross domestic product divided by midyear population. GDP is the

sum of gross value added by all resident producers in the economy plus any product
taxes and minus any subsidies not included in the value of the products. It is calcu-
lated without making deductions for depreciation of fabricated assets or for deple-
tion and degradation of natural resources. Data are in constant US dollars from the
year 2000.
GDSR: The gross domestic savings rate is calculated as GDP less final con-
sumption expenditure (total consumption), as a percent of GDP. These numbers
(as reported by the World Bank) are from World Bank national accounts data and
OECD National Accounts data.
REAL INTEREST RATE is the lending interest rate of a country at time t adjusted
for inflation as measured by the GDP deflator. The source for these numbers (as
reported by the World Bank) is the International Monetary Fund, International
Financial Statistics, using World Bank data on the GDP deflator.
OLDt and YOUNGt are the percent of the population that are older than 65 and
younger than 15 in year t.
D. Variables in the SHARE
HHNETWORTH: A household’s net worth in the SHARE “HHNetWorth” is

attempt to measure all real assets net of any debts on them. It is equal to the esti-
mated value of a household’s: main residence, real estate other than the main resi-
dence, businesses, cars, bank accounts, bonds, stocks, mutual funds, life insurance,
minus mortgage and other debt, in 2005 Euros, as collected by the OECD.
AVERDISINC: A country’s average disposable income is per head, with PPP and
in 2005 Euros, as collected by the OECD.
SMOKED: This codes whether an individual reports: “Have you ever smoked
cigarettes, cigars, cigarillos or a pipe daily for a period of at least one year?”
PHYSICALLY ACTIVE: Physical inactivity is defined as “never or almost never
engaging in neither moderate nor vigorous physical activity.” Being physically
active is not being inactive.
Table A4 —Data Summary: DHS, Household Level
Smoke 4.83
Obesity (BMI > 30) 23.1
Use any form of contraception 24.7
Used a condom last sexual encounter 8.85
Sex (male) 29.9
Education: Primary school 22.1
Education: Secondary school 29.2
Education: Higher 6.11
Mean SD
Age (years) 29.3 10.4
Household size 6.74 4.19
Table A5—Data Summary: OECD Country-Year Level
Mean SD
GDSRt 24.2% 5.90%
Strong FTR 64.6% 46.6%
PCGDPt(US$) $22,096 $9,259
PCGDPt−1/PCGDPt 0.978 0.030
CAGRt 9.69% 9.90%
Oldt 13.1% 3.48%
Youngt 20.2% 4.94%
Social Securityt 0.743 0.233
Distance from equator 3.214 0.666
Legal origins (number of countries):

UK (7), French (12), German (11), Scandinavian (5)
OBESITY: This is defined as a body-mass index of 30 or greater.

WALKING SPEED: This was measured only among individuals aged 76 years
and older. Walking speed was averaged over two tests, as measured in meters per
second, down a hallway at least 10 meters long.
GRIP STRENGTH: Grip strength is measured with a dynamometer at the inter-
view (in kg).
PEAK FLOW: Peak expiratory flow measures a person’s maximum exhalation
air-flow, as measured with a peak-flow meter (in L/min).
E. Variables in the DHS
SMOKE: This codes whether an individual reports currently smoking.

OBESITY: This is defined as a body-mass index of 30 or greater.
CONTRACEPTION: This codes whether the respondent reports currently using
any form of contraception.
CONDOM: This codes whether the respondent reports having used a condom
during their last sexual encounter.
WEALTH and EDUCATION are coded categorically in the DHS. Wealth is

reported as a five-level wealth index, while education is coded by highest level of
education attained.
F. OECD Variables
All GDP-based measures are computed using the expenditure method, with con-
stant PPPs using US dollars from the OECD base year (2000).
CAGR is the average growth rate of the country from 1993 to 2009 (the earliest
date for which data is available for all countries).
OLDt and YOUNGt are the percent of the population that are older than 65 and
younger than 15 in year t.
SOC SECtis the per capita fraction of GDP spent by a country in year t on social
security payments, divided by the fraction of the population that are over 65. These
payments include expenditures on disability, old age, and survivors benefits.
DIST FROM EQUATOR is the distance in thousands of miles between a coun-
try’s capital and the equator.
Appendix B: Measures of Future-Time Reference
A. Methods for the Online Measures
Selection of Languages to Cover.—As a first proxy for languages which are well
represented on the web, I look at the set of languages that Google allows a web
search to be restricted to, or which are covered by Google Translate (whose main
function is translating websites). I exclude languages from this list which are either
synthetic (Esperanto) or are not spoken by significant numbers of people as their
first language (Latin). I then conduct a Google search for variants of the phrase
“weather forecast,” in each of these languages, restricted to results in that language.
So for example, I conduct a Google search for the terms “wettervorhersage,” “wet-
terprognose,” and “wetterberichte,” restricted to websites in German.
Gathering Texts.—From here, research assistants and I gathered websites indexed

within the first five pages (or 60 results) returned by Google.40 We identified those
websites which contained full-sentence weather forecasts, as opposed to forecasts
expressed pictorially (sun and cloud icons), or as short phrases (“Friday, high of
62”). For several sparsely spoken languages we could find no such websites, and
excluded those languages from analysis. This selection methodology resulted in a
set of 39 languages for analysis, with the vast majority represented by three or more
websites.
For each of these 39 languages, we then scraped the web over a period of three
months (5–2012 through 7–2012), collecting forecasts from the websites we had
identified. Restricting analysis to sentences from these scrapings that refer solely
40
I want to thank Yale undergraduates Jane Bang and Ryan Caro, for their invaluable research assistance in
assembling these data, and Yale linguistics doctoral candidate Nicole Palffy-Muhoray for invaluable feedback on
sensible measures of future-time reference intensity.
to future events (some sentences discuss past weather patterns, or general seasonal
patterns), resulted in roughly 46 sentences per language to analyze, for an average
of 58.4 verbs.
Computing Measures.—In each language, we compiled the set of grammatical

markers which linguists agree are future-time markers. Then, for each of these 39
languages, we computed two measures of future-time reference intensity. “Verb
ratio” counts the number of verbs which are grammatically future-marked, divided
by the total number of future-referring verbs. In other words, in online weather
forecasts in a language, what share of verbs about future weather are marked as
future-referring? Similarly, “sentence ratio” asks: what share of sentences regarding
future weather contain a grammatical future marker? In some languages (Arabic,
for example), a sentence with multiple verbs will often mark only the first as future-
regarding. Grammatical differences like this produce variation between verb and
sentence ratios. The results of this exercise are summarized in Table B1.
Regressions with Online Language Measures.—While the set of languages cod-

able in this way is limited to those which are well represented on the searchable
Internet, it is extensive enough that both the OECD and SHARE results I report can
be run substituting either ratio instead of the binary weak- versus strong-FTR mea-
sure. Both measures produce results that are nearly identical (both quantitatively
and statistically) to the results I report in this paper. Please see the online Appendix
for the results of these regressions.
B. Methods for Extending the EUROTYP
The analyses of Thieroff (2000) suggest that the tendency to mark prediction-
based FTR maps more generally onto whether “future time reference can be referred
to with unmarked form (the present).” Dahl also finds that weak-FTR corresponds
strongly with a language’s general tendency to require FTR, and suggests that
“whether FTR is overtly and obligatorily marked in prediction-based sentences can
be used as one of the major criteria for whether it is grammaticalized in a language
or not” (Dahl 2000). These analyses motivate my decision to use weak-FTR as a
proxy for the general treatment of future time in a language.
Most analyses in this paper study languages directly analyzed by the EUROTYP
Theme Group. In those regressions, weak-FTR languages are the set of languages
Dahl calls “futureless” languages and Thieroff (2000) calls “weakly-grammatical-
ized future” languages. Some regressions analyze the World-Values Survey, whose
participants speak many non-European languages not analyzed by either Dahl or
Thieroff.
To extend their characterization to this broader set, I rely on several other cross-
linguistic analyses that have studied how languages mark future time (most nota-
bly Bybee, Perkins, and Pagliuca 1994; Cyffer, Ebermann, and Ziegelmeyer 2009;
Dahl 1985; Dahl and Kós-Dienes 1984; and Nurse 2008). Most importantly, several
African countries are well represented in the WVS and have several national lan-
guages. Given their potential importance for within-country identification, I code
these languages only when both a cross-linguistic study and a language specific
Table B1—Languages and Online FTR Ratios
Verb ratio Sentence ratio Strong

Language (percent) (percent) FTR
Azerbaijani 100.0 100.0 Strong
Basque 98.4 100.0 Strong
Catalan 100.0 100.0 Strong
Greek 97.4 100.0 Strong
Hebrew 100.0 100.0 Strong
Irish 100.0 100.0 Strong
Korean 82.2 100.0 Strong
French 95.8 97.6 Strong
Albanian 98.4 97.5 Strong
Lithuanian 93.2 97.2 Strong
Belarusian 93.5 96.4 Strong
Bulgarian 93.8 95.5 Strong
Romanian 96.1 95.1 Strong
Slovenian 81.5 94.4 Strong
English (UK) 88.1 92.9 Strong
Italian 90.0 92.9 Strong
English (US) 76.9 87.5 Strong
Maltese 86.4 82.4 Strong
Portuguese (EU) 85.0 81.3 Strong
Russian 72.2 80.8 Strong
Croatian 78.6 80.0 Strong
Spanish 71.6 74.1 Strong
Turkish 55.8 66.7 Strong
Vietnamese 59.6 66.7 Strong
Latvian 58.3 55.2 Strong
Czech 46.4 54.5 Strong
Arabic 41.7 52.9 Strong
Polish 28.2 34.4 Strong
Hungarian 25.0 32.3 Strong
Norwegian 15.3 20.9 Weak
Danish 10.0 12.5 Weak
Swedish 4.9 6.3 Weak
Chinese 0.0 0.0 Weak
Dutch 0.0 0.0 Weak
Estonian 0.0 0.0 Weak
Finnish 0.0 0.0 Weak
German 0.0 0.0 Weak
Japanese 0.0 0.0 Weak
Portuguese (BR) 0.0 0.0 Weak
reference grammar agree on a language’s FTR structure. Most important were Adu-
Amankwah (2003) for Akan; Olawsky (1999) and Lehr, Redden, and Balima (1966)
for Dagbani and Moore; Newman (2000) for Hausa; Carrell (1970), Emenanjo
(1978), Ndimele (2009), and Uwalaka (1997) for Igbo; Bentley (1887) for Kongo
and Awobuluyi (1978); and de Gaye and Beecroft (1964) for Yoruba.
I have attempted to be as conservative as possible in extending my weak-FTR
coding to languages not covered by the EUROTYP. Several rules applied, which I
will describe.
A Conservative Binary Coding.—For several languages, multiple sources suggest

that language does not grammatically mark future events. For example, Chinese
grammars uniformly find no grammatical marking of future events, and amply
available texts demonstrate that Chinese speakers grammaticalize the future exactly
like they grammaticalize the present and past. Similarly, Finnish and Estonian are
languages in Europe which stand out as lacking FTR grammaticalization. Among

African languages, Nurse (2008) analyzes a set of more than 200 Bantu languages,
and finds that roughly 9 percent of those languages grammaticalize a non-past
category, (that is, have no discretely grammaticalized future). Kongo is a notable
example. Dahl (1985) studies 64 languages from every major language group in the
world, and finds 14 that show no evidence of grammaticalized FTR.
• If a language is described as lacking any grammaticalized FTR by several inde-
pendent sources, I code it as weak-FTR.
Similarly, in these cross-linguistic studies, several languages stand out as having
particularly heavily-grammaticalized FTR. For example, Nurse (2008) notes several
languages in which not only are future events grammaticalized with a dedicated pre-
fix or suffix, but often posses finer obligatory distinctions, like a hodiernal (before
dawn tomorrow) future. Studies like Dahl (1985) allow me to add a quantitative
component to this kind of comparison. In his surveys of native speakers, there are
languages (like Georgian), which both posses a dedicated inflectional future, and
whose speakers use this future in nearly every sentence with FTR.
• If multiple sources describe a language in this way, I code it as strong-FTR.
Note that this process is conservative, as there are several languages studied in
the EUROTYP which these rules would not have been able to classify. For example,
Swedish was classified as weak-FTR (“futureless”) by both Dahl and Theiroff, but
is a language which I would not have been able to classify. Please see the online
Appendix for a full table of all languages included in this study and their coding.
Robustness.—An alternative process would be a web-data scraping exercise like

the one I describe above, which would result in a continuous (though imperfect in
other ways) measure. This suggests two natural robustness checks. First, I could
have used the continuous web measure in my regressions. Second, I could have
continued to use a binary classification, but tested the sensitivity of my results to
moving the cut-off which defines the binary classification along the continuous
web measure.
As of now, the small sample of languages that I can analyze using my web-
weather-forecast methodology is too small to be useful for WVS regressions. In my
regressions in both the SHARE and the OECD data though, both of these robustness
checks are possible, and produce results which are both quantitatively and statisti-
cally identical to the ones I report in thin paper. This gives me confidence that the
procedure I adopted to expand the set of weak- and strong-FTR coded countries did
not systematically bias my results. Please see the online Appendix for the results of
these robustness checks.
References
Acemoglu, Daron, and James A. Robinson. 2012. Why Nations Fail: The Origins of Power, Prosperity,
and Poverty. New York: Crown.
Adu-Amankwah, David. 2003. Asante-Twi Learners’ Reference Grammar, Madison, WI: National
African Language Resource Center Press.
Arrow, Kenneth J., and David Levhari. 1969. “Uniqueness of the Internal Rate of Return with Variable
Life of Investment.” Economic Journal 79 (315): 560–66.
Awobuluyi, Oladele. 1978. Essentials of Yoruba Grammar, Ibadan, Nigeria: Oxford University Press.
Barro, Robert J., and Glenn M. MacDonald. 1979. “Social Security and Consumer Spending in an
International Cross Section.” Journal of Public Economics 11 (3): 275–89.
Börsch-Supan, Axel, and Hendrik Jürges, eds. 2005. The Survey of Health, Aging and Retirement in
Europe—Methodology. Mannheim: Research Institute for the Economics of Aging.
Bentley, Holman. 1887. Dictionary and Grammar of the Kongo Language. London: Trubner.
Bittner, Maria. 2005. “Future Discourse in a Tenseless Language.” Journal of Semantics 22 (4):
339–87.
Brown, Roger H., and Eric H. Lenneberg. 1954. “A Study in Language and Cognition.” Journal of
Abnormal and Social Psychology 49 (3): 454–62.
Burbidge, John B., Lonnie Magee, and A. Leslie Robb. 1988. “Alternative Transformations to Han-
dle Extreme Values of the Dependent Variable.” Journal of the American Statistical Association
83 (401): 123–27.
Bybee, Joan, Revere Perkins, and William Pagliuca. 1994. The Evolution of Grammar, Chicago: Uni-
versity of Chicago Press.
Carrell, Patricia L. 1970. A Transformational Grammar of Igbo. Cambridge: Cambridge University Press.
Chamberlain, Gary. 1980. “Analysis of Covariance with Qualitative Data.” Review of Economic Stud-
ies 47 (1): 225–38.
Chen, M. Keith. 2013. “The Effect of Language on Economic Behavior: Evidence from Savings Rates,
Health Behaviors, and Retirement Assets: Dataset.” American Economic Review. http://dx.doi.
org/10.1257/aer.103.2.690.
Chomsky, Noam. 1957. Syntactic Structures. Paris: Mouton and Co.
Comrie, Bernard. 1985. Tense. Cambridge: Cambridge University Press.
Copley, Bridget. 2009. The Semantics of the Future. New York: Routledge.
Croft, William. 2003. Typology and Universals. second ed. Cambridge: Cambridge University Press.
Cyffer, Norbert, Erwin Ebermann, and Georg Ziegelmeyer. 2009. Negation Patterns in West African
Languages and Beyond. Amsterdam: John Benjamins Publishing Company.
Dahl, Östen, and Dora Kós-Dienes. 1984. Selected Working Papers from the Tense-Mood-Aspect Proj-
ect. Stockholm: Institute of Linguistics, University of Stockholm.
Dahl, Östen. 1985. Tense and Aspect Systems. Oxford: Basil Blackwell.
Dahl, Östen. 2000. “The Grammar of Future Time Reference in European Languages.” In Tense and
Aspect in the Languages of Europe, edited by Östen Dahl, 309–28. Berlin: Mouton de Gruyter.
Dahl, Östen, and Viveka Velupillai. 2011. “The Future Tense.” In The World Atlas of Language Struc-
tures Online, edited by Matthew Dryer and Martin Haspelmath. Munich: Max Planck Digital
Library. http://wals.info/chapter/67 (accessed June 1, 2011).
Dancygier, Barbara, and Eve Sweetser. 2009. Mental Spaces in Grammar: Conditional Constructions.
Cambridge: Cambridge University Press.
de Gaye, J.A. and W.S. Beecroft. 1964. Yoruba Grammar, London: Routledge.
Diamond, Jared. 2005. Guns, Germs, and Steel: The Fates of Human Societies. London: W. W. Nor-
ton and Co.
Emenanjo, E. Nolue. 1978. Elements of Modern Igbo Grammar, Ibadan, Nigeria: Oxford University Press.
Feldstein, Martin. 1980. “International Differences in Social Security and Saving.” Journal of Public
Economics 14 (2): 225–44.
Franklin, A., G. V. Drivonikou, L. Bevis, I. R. L. Davies, P. Kay, and T. Regier. 2008. “Categorical Per-
ception of Color is Lateralized to the Right Hemisphere in Infants, but to the Left Hemisphere in
Adults.” Proceedings of the National Academy of Sciences 105 (9): 3221–25.
Frederick, Shane, George Loewenstein, and Ted O’Donoghue. 2002. “Time Discounting and Time
Preference: A Critical Review.” Journal of Economic Literature 40 (2): 351–401.
Harner, Lorraine. 1981. “Children Talk about the Time and Aspect of Actions.” Child Development
52 (2): 498–506.
Hicks, John. 1973. Capital and Time. Oxford: Clarendon Press.
Iatridou, Sabine. 2000. “The Grammatical Ingredients of Counterfactuality.” Linguistic Inquiry 31 (2):
231–70.
Jakobson, Roman, and Morris Halle. 1956. Fundamentals of Language. The Hague: Mouton and Company.
James, Deborah. 1982. “Past Tense and the Hypothetical: A Cross-Linguistic Study.” Studies in Lan-
guage 6 (3): 375–403.
Kacelnik, Alex, and Melissa Bateson. 1996. “Risk Theories—The Effects of Variance on Foraging
Decisions.” American Zoologist 36 (4): 402–34.
La Porta, Rafael, Florencio Lopez-de-Silanes, and Andrei Shleifer. 2008. “The Economic Conse-
quences of Legal Origins.” Journal of Economic Literature 46 (2): 285–332.
La Porta, Rafael, Florencio Lopez-de-Silanes, Andrei Shleifer, and Robert W. Vishny. 1997. “Legal
Determinants of External Finance.” Journal of Finance 52 (3): 1131–50.
Landes, David S. 1998. The Wealth and Poverty of Nations: Why Some are so Rich and Some so Poor.
New York: W. W. Norton and Co.
Lehr, Marianne, James E. Redden, and Adam Balima. 1966. More, a Basic Course. Washington, DC:
United States State Department, Foreign Service Institute.
Lewis, M. Paul, ed. 2009. “Ethnologue: Languages of the World.” 16th ed. Dallas, TX: SIL Interna-
tional. Online version: http://www.ethnologue.com/ (accessed August 6, 2012).
MacLaury, Robert E., Gordon W. Hewes, Paul R. Kinnear, J. B. Deregowski, William R. Merrifield,
B. A. C. Saunders, James Stanlaw, Christina Toren, J. Van Brakel, Roger W. Wescott. 1992. “From
Brightness to Hue: An Explanatory Model of Color-Category Evolution.” Current Anthropology
33 (2): 137–86.
Marschak, Jacob. 1965. “Economics of Language.” Behavioral Science 10 (2): 135–40.
Mezhevich, Ilana. 2008. “A Time-Relational Approach to Tense and Mood.” In Proceedings of the 27th
West Coast Conference on Formal Linguistics, edited by Natasha Abner and Jason Bishop, 326–34.
Somerville, MA: Cascadilla Press.
Modigliani, Franco. 1986. “Life Cycle, Individual Thrift, and the Wealth of Nations.” American Eco-
nomic Review 76 (3): 297–313.
Ndimele, Ozo-mekuri. 2009. “Negation Markings in Igbo.” In Negation Patterns in West African Lan-
guages and Beyond, edited by Norbert Cyffer, Erwin Ebermann, and Georg Ziegelmeyer, 121–37.
Amsterdam: John Benjamins Publishing Co.
Newman, Paul. 2000. The Hausa Language: An Encyclopedic Reference Grammar, New Haven: Yale
University Press.
Nurse, Derek. 2008. Tense and Aspect in Bantu. New York: Oxford University Press.
Olawsky, Knut J. 1999. Aspects of Dagbani Grammar: With Special Emphasis on Phonology and Mor-
phology. Munich: LINCOM Europa.
Organisation for Economic Co-operation and Development (OECD). 2011. “National Accounts of
OECD Countries.” Washington, DC: OECD.
Pinker, Steven. 1994. The Language Instinct: How the Mind Creates Language. New York: William
Morrow and Company.
Poterba, James M., ed. 1994. International Comparisons of Household Saving. National Bureau of
Economic Research Project Report series. Chicago: University of Chicago Press.
Poterba, James M., Steven F. Venti, and David A. Wise. 1995. “Do 401(k) Contributions Crowd Out
Other Personal Saving?” Journal of Public Economics 58 (1): 1–32.
Redelmeier, Daniel A., and Daniel N. Heller. 1993. “Time Preference in Medical Decision Making and
Cost-Effectiveness Analysis.” Medical Decision Making 13 (3): 212–17.
Rubinstein, Ariel. 2000. Economics of Language. Churchill Lectures in Economic Theory. New York:
Cambridge University Press.
Samuelson, P. A. 1937. “Some Aspects of the Pure Theory of Capital.” Quarterly Journal of Econom-
ics 51: 469–96.
Saussure, Ferdinand de. 1916. Course in General Linguistics. Edited by Charles Bally and Albert
Sechehaye. Translated by Roy Harris. La Salle, IL: Open Court, 1983.
Scholz, Barbara C., Francis Jeffry Pelletier, and Geoffrey K. Pullum. 2011. “Philosophy of Linguis-
tics.” In The Stanford Encyclopedia of Philosophy, Winter 2011 Edition. Edited by Edward N. Zalta.
http://plato.stanford.edu/archives/win2011/entries/linguistics/.
Schiffrin, Deborah. 1981. “Tense Variation in Narrative.” Language 57 (1): 45–62.
Szagun, Gisela. 1978. “On the Frequency of Use of Tenses in English and German Children’s Sponta-
neous Speech.” Child Development 49 (3): 898–901.
Thieroff, Rolf. 2000. “On the Areal Distribution of Tense-Aspect Categories in Europe.” In Tense and
Aspect in the Languages of Europe, edited by Östen Dahl, 309–28. Berlin: Mouton de Gruyter.
Uwalaka, Mary Angela. 1997. Igbo GramMarch Ibadan, Nigeria: The Pen Services.
Winawer, Jonathan, Nathan Witthoft, Michael Frank, Lisa Wu, Alex Wade, and Lera Boroditsky.
2007. “Russian Blues Reveal Effects of Language on Color Discrimination.” Proceedings of the
National Academy of Sciences 104 (19): 7780–85.
Wittgenstein, Ludwig. 1922. Tractatus Logico-Philosophicus. London: Routledge.
Woolley, Frances. 2011. “A Rant on Inverse Hyperbolic Sine Transformations.” Worthwhile Canadian
Initiative (blog). http://worthwhile.typepad.com/worthwhile_canadian_initi/2011/07/a-rant-on-
inverse-hyperbolic-sine-transformations.html.
World Values Survey Association. 2009. “World Values Survey 1981–2008, Official Aggregate.”
v.20090901. ASEP/JDS, Madrid. www.worldvaluessurvey.org.
Whorf, Benjamin Lee. 1956. Language, Thought, and Reality: Selected Writings of Benjamin Lee
Whorf. Edited by John B. Carroll. Cambridge, MA: MIT Press.

The Effect of Language On Economic Behavior

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Effect of Language On Economic Behavior

Uploaded by

Copyright:

Available Formats

American Economic Review 2013, 103(2): 690–731

The Effect of Language on Economic Behavior:

Languages differ in whether or not they require speakers to grammatically mark

future-oriented in numerous monetary and non-monetary behaviors. Weak-FTR

I. Language and Future-Time Reference

“… nineteen verb-extending suffixes (sixteen transitivity preserving … ,

(1) W​ oˇ ​ qù t​ ​

(2) a. Tänään on kylmää

Today be.prs cold

(3) a. Il fait froid aujourd’hui

b. Il fera froid demain

English is a notable outlier in Europe; in all other Germanic languages grammatical

(4) a. Morgen regnet es

b. Morgen ist es kalt

II. A Linguistic-Savings Hypothesis

My linguist-savings hypothesis stems naturally from two different types of mech-

A. Mechanism One: Obligatory Distinctions Bias Beliefs

“There is a famous story of President Abraham Lincoln, taking a vote in

“I wish I had a car (right now).”

“I wish I had had a car (when I was a student).”

(6) ​ ​ < ​δS​​, then ∫ ​e−​

​ ​S​(t) would first-order stochastically dominate F

B. Mechanism Two: Linguistic Distinctions Lead to More Precise Beliefs

III. Data and Methods

In all of the regressions to follow, the independent variable of main interest is

granular obligatory d­ iscretization of future time) that is central to the mechanism I

Online-Text Based Codings.—To my knowledge, the EUROTYP project is the

C. Savings Regressions in the WVS

D. Retirement Assets in the SHARE

E. Health Behaviors in the SHARE and DHS

F. National Savings Rates

Finally, I study the relationship between language and savings in a cross-country

rate of GDP as well as a number of other country demographics. To this regression

A. Language and Savings in the World Values Survey

My first set of regressions examines the savings behavior of individuals. I examine

In equation (9), the main variable of interest StrongFTR is a binary-coded charac-

Table 1—An Individual Saved This Year (WVS, Cross-Country Analysis)

Observations 152,056 149,350 140,498 140,498 134,535 134,535

Table 2—WVS Cross-Country Analysis by Continent and PCGDP

Regression restricted by PCGDP:

of household savings in Poterba (1994). Empirical estimates of equation (10) are

Table 3—An Individual Saved This Year (WVS, Within-Country Analysis)

Observations 152,056 64,017 64,017 24,933 23,615 23,615

Table 4—WVS Countries with Large Within-Country FTR Differences

Table 5—Additional Within-Country Control Regressions in the WVS

Observations 21,834 1,781 23,615 13,245 13,245

Figure 1. Between-Country Savings and Language in the WVS

Table 6—Household Retirement Assets (SHARE)

(1) (2) (3) (4) (5)

Observations 39,665 39,665 39,665 39,350 5,937

of language. Regressions 4 and 5 include fixed effects for religious denomination

B. Language and Retirement Assets in Europe

IHS(x) = log​ x + (​x​ 2​ + 1​)​2 ​ ​. ( _ 1

In equation (11) the dependant variable is IHS​( _

Table 7—Household Retirement Assets (SHARE, BE, and CH)

(1) (2) (3) (4) (5) (6)

Observations 4,410 4,409 1,553 1,553 148 148

C. Language and Health in the SHARE

Table 8—Health Behaviors and Measures of Health (SHARE)

Observations 15,750 9,135 11,958 6,038 51,571 26,836

of health. While there appears to be no effect on walking speed, speaking a strong-

D. Language and Health in the DHS

Table 9—Health Behaviors in Developing Countries (DHS)

(1) W   oˇ  qù t   

(6)  < δS, then ∫ e−

S(t) would first-order stochastically dominate F

granular obligatory d iscretization of future time) that is central to the mechanism I

IHS(x) = log x + (x  2 + 1)2   . (  _ 1

In equation (11) the dependant variable is IHS( _